CVAT Academy

Data annotation is the process of structuring information so that a computer can "understand" it. It is the foundation for training machine learning models and artificial intelligence (AI) systems. Without properly annotated data, it is impossible to build a high-quality model.Data annotation is done by annotators. An annotator is a person who manually labels data, indicating what is shown in an image, said in audio, or written in text. Their task is not just to "draw," but to turn raw data into training material for AI.Here is how the process of annotating a car using a bounding box looks. In this way, we define the location of the car in the image and label it as “Car.”For AI, this means: "This is what a car looks like — learn to recognize such objects on your own."Where is data annotation used?Data annotation is used in many fields where artificial intelligence and machine learning are involved.Automotive & Autonomous DrivingAnnotation is the basis for teaching smart machines to understand what’s happening on the road.What is annotated:Pedestrians, cars, road signs, traffic lightsLane markings, road structureDanger zones and obstaclesExamples:Tesla, Waymo — use annotated video data to train autopilot systems.MedicineAnnotation can save lives — it helps AI learn to detect diseases in medical scans.What is annotated:Tumors, blood clots, fractures in MRI, CT, and X-raysOrgan contours (lungs, heart, brain)Cells, viruses, bacteria in microscope imagesExamples:Cancer detection, tuberculosis, COVID-19, automatic blood cell counting.Retail & StoresAI helps optimize sales by analyzing customer behavior.What is annotated:Products on shelves, their positioningVisitors: how many people, where they look, where they standEmotions, age, gender (for personalized advertising)Examples:Cameras in stores record video used to train models to recognize inventory and customer activity.Satellite & Drone ImageryAnnotation is used in agriculture, ecology, and defense.What is annotated:Fields, roads, buildings, bodies of waterTrees, animals, crops, fires, destructionExamples:Google Maps, agricultural monitoring, illegal deforestation detection.Security & SurveillanceAI recognizes suspicious behavior or identities in a video feed.What is annotated:Faces, human figures, suspicious actionsAccess to restricted zonesExamples:Monitoring in train stations, stadiums, airports.Education & AI ResearchUniversities and research labs create datasets to train and test new models.What is annotated:Images, text, audio, and video for open science projectsExamples:COCO, ImageNet, Cityscapes — world-famous datasets created by manual annotation.What problem does annotation solve?Annotation is not just adding labels to data — it solves the core problem of making data understandable for machine learning.The main goal of annotation is to prepare raw data so that a neural network can learn from it and make intelligent predictions.In practice, machines don't understand images like humans — they only see pixels.Annotation explains: “This is a pedestrian,” “Here’s a dog,” “This is a field boundary.”This information becomes the “correct answer” for training the model.Without annotation, a model would:Have no idea what is right or wrongLearn “blindly” from noisy, unstructured dataMake unacceptable errors in real-world tasks (especially in medicine or transportation)Examples of problems annotation solves:A car must recognize pedestrians → people must be annotated in imagesMedical AI for diagnosis → tumors must be annotated in scansAI for shelf inventory → every product must be labeledIn summary: annotation turns data into knowledge — the foundation for artificial intelligence.Main Types of Annotation in CVATWhen working with CVAT (Computer Vision Annotation Tool), you'll encounter various types of annotation. Here's a quick overview:Classification (What is in the image?)Definition:The model sees the entire image and determines what is depicted, without specifying object locations.Process:The annotator selects an appropriate label for the whole image using tags.Example:Key point:No location specifiedOnly the meaning of the image is capturedObject Detection (What and where is it?)Definition:Now it’s important not only what is in the image, but also where the objects are located.Rectangles (bounding boxes) are drawn around objects, each labeled accordingly.Process:Draw a box around each object and label it (e.g., “car,” “person,” “bike”).Example:Key point:Object count and location are visibleSegmentation (Precise shape of the object)Definition:More advanced than bounding boxes — draw the exact shape of the object to show its outline in detail.Process:Draw the object’s boundary using polygons or masks.Example:Key point:Maximum precisionAnnotation of an object with background exclusionTracking (Follow the object across video)Definition:In video, it’s important not only to find an object but to follow it across frames.The goal is to understand that it’s the same object moving through the scene.Process:Draw a box or contour in each frameAssign an ID to track the same object through timeExample:Key point:Object is tracked over timeEnables analysis of movement, speed, directionSkeleton Annotation (Keypoints and body pose)Definition:This type of annotation focuses on identifying the precise position of specific keypoints on an object — most often a human or animal. These keypoints are then connected to form a skeleton that represents the pose or posture.Process:The annotator marks each keypoint (e.g., head, shoulders, elbows, knees, etc.) and connects them according to a predefined skeleton structure.Each keypoint can also have visibility flags (e.g., visible, occluded, not labeled).Example:Key point:Allows detailed analysis of body movements and posturesEssential in sports analytics, gesture recognition, animation, healthcare, and safety monitoringOften used together with tracking to follow poses across video frames3D Annotation (Depth and spatial position of objects)Definition:3D annotation allows annotators to mark objects in three-dimensional space, adding depth information to traditional 2D labeling. This is essential in applications like autonomous driving, robotics, and augmented reality.Process:Use cuboids (3D boxes) to surround the object. Each cuboid includes not only width and height, but also depth and orientation in space. The annotator adjusts the cuboid’s perspective to align it with the object's shape and position in the scene.‍Example:Key point:Captures real-world spatial orientation and sizeCrucial for depth-aware modelsUsed with multi-view camera systems or LiDAR dataAdds complexity, but provides richer geometric context‍‍‍Roles in the Annotation ProcessAnnotatorThe main worker who manually annotates data according to the instructions.Responsibilities:Study the guidelines carefully before startingAnnotate accurately: use the correct tools and name object classes properlyMeet deadlines and required volumeAsk the team lead when in doubtValidatorChecks completed annotation against instructions and identifies errors. Often works after annotation is done.Responsibilities:Assess quality: accuracy of contours, correct classes, completenessMark and fix errors (if allowed)Request revision and provide feedbackEnsure project quality standards are metTeam LeadManages and oversees the annotation process. Connects the team with the client.Responsibilities:Create and explain the guidelinesTrain team members, answer questionsDistribute tasks among annotators and validatorsMonitor deadlines and progressPerform spot checksCommunicate with the client and deliver resultsAnnotation AccuracyAnnotation accuracy is a critical quality metric. Well-annotated data directly impacts the effectiveness of machine learning models. Even small errors can degrade results.Precise object positioning — the contours or bounding boxes must closely follow the shape of the object without excessive background or omissions. If capturing more background is acceptable in the project, such tolerances must be clearly described in the specification.In this example, the person on the right is inaccurately annotated. Red arrows point to critical issues such as large gaps and clipped regions. In most projects, this type of annotation would be considered invalid.Correct class assignment — each object must be annotated strictly according to the guidelines.In the example on the right, the object has been assigned an incorrect label. The correct label should be "bicycle" rather than "motorcycle." Such mistakes significantly reduce annotation quality and should be avoided.Annotation completeness — all required objects in the image must be annotated, none should be missing.In this example, one person was not annotated. An annotation task must always be complete, with all required objects properly labeled.Guideline compliance — each tool (bounding box, polygon, line, etc.) must be used appropriately, as specified in the project documentation. All nuances described in the specification must be taken into account.In this example, the goose indicated by the red arrow is annotated using a segmentation mask, while the others are annotated with bounding boxes. Annotation must be consistent and strictly follow the project specification. If the specification requires the use of bounding boxes only, using any other annotation tools is not allowed.ConclusionData annotation is not just a technical task — it's a foundational step in building modern AI systems. The accuracy of annotation determines how “smart” and reliable a model will be when solving real-world problems.Annotation requires attention to detail, full understanding of the guidelines, and responsibility for the final result. The annotator transforms raw data into training material; the validator ensures precision; the team lead coordinates everything to run smoothly.Even the most advanced AI algorithms are useless without high-quality annotated data. That’s why annotation isn’t a side job — it’s the core of a successful AI project.

When we hear about AI, we often imagine smart machines, autopilots, voice assistants, or diagnostic tools in medicine. But behind each of these “intelligent” solutions are thousands of hours of human labor. One of the key roles in this process is the data annotator.A data annotator is not just someone who draws boxes or polygons. They are a specialist responsible for the quality of input data used to train intelligent systems. Every annotation created becomes part of a training dataset for neural networks. The efficiency of AI systems in the future directly depends on the annotator's precision, attention to detail, and understanding of the instructions.Key Responsibilities of an AnnotatorAnnotation is not just a mechanical task — it requires careful, responsible work, understanding of project goals, precision, and discipline.Studying the SpecificationFor an annotator, the specification is not a formality but a primary work tool. All annotation decisions are based on it. Before starting a project, the annotator must thoroughly and carefully study the specification. This is essential not only to understand what and how to annotate but also to avoid mistakes, rework, and time loss later.In large and complex projects, the specification can be extensive — dozens of pages. Such documents are impossible to memorize after one reading. So during the first days of work (and sometimes longer), annotators must regularly refer back to the instructions — check details during annotation and clarify edge cases.A good annotator doesn’t rely on memory or guesswork. They work carefully and systematically:Always keep the specification handy,Refer back to it at the slightest doubt,Aren’t afraid to ask questions when something is unclear.Ignoring the specification or “eyeballing” leads to errors, decreases data quality, and causes re-annotation.Precise and Neat AnnotationOne of the annotator’s main goals is to annotate precisely, carefully, and neatly. Even when following instructions, poor quality shapes, labels, or borders can render data useless. Each annotation should:Match object boundaries exactly,Be logically shaped and structured,Contain no missing or extra elements,Appear clean and visually accurate.Annotation isn’t just about “tracing” an object — it’s about doing it with the required accuracy. Low precision causes information loss, while excessive detail introduces noise and leads to model errors.Accuracy and speed: balancing project requirementsThe required level of accuracy varies by project. Some demand pixel-perfect annotation with flawless polygons and zero gaps. Others allow rougher annotation, with minor object clipping, small gaps, or angular shapes.All acceptable tolerances and accuracy levels should be clearly defined in the specification. If they aren’t, the annotator must ask the team lead. This prevents unnecessary rework and misinterpretation.Accuracy directly affects annotation speed:Higher accuracy means more time spent on details, shape control, and edge alignment.If the project allows looser annotation, the annotator can work faster while staying within guidelines.This is critical for meeting deadlines. If an annotator spends time on ultra-detailed annotation in a project where it’s unnecessary, they risk missing deadlines. This “over-precision” harms not just speed but overall team performance.Work ApproachEven when annotating a large batch of similar objects, it’s vital to stay focused. Common errors occur when the work becomes routine:Crooked polygons,Misaligned boxes,Incomplete lines,Cropped object edges.A professional annotator regularly self-checks:Takes short breaks to review work with a fresh eye,Revisits earlier images,Compares results with instruction examples.Neatness isn’t just a good habit — it’s a quality requirement. Careful annotation prevents rework, saves team resources, and impacts overall project success.Meeting DeadlinesMeeting deadlines isn’t just a formality — it’s a vital part of the annotator’s job. Even perfect annotations lose value if delivered too late, disrupting the project schedule.Annotation projects often involve dozens of people, and any delay affects the entire chain: from validation to final delivery. Annotators must understand their role in this system and avoid delays unless absolutely necessary.Factors Influencing TimelinessSelf-assessment: Taking on too many tasks can lead to delays. It’s better to accept fewer and complete them on time.Project complexity: More detailed annotation requires more time per item. Don’t expect the same pace as in simpler projects.Instruction availability: For new projects, time must be allocated for learning requirements and examples. The first tasks are always slower.Work organization: Regular work, time management, and reducing distractions help prevent backlog and ensure on-time delivery.Risks of Missing DeadlinesTeam-wide delays: One annotator’s delay postpones validation and risks the final deadline.Loss of trust: Repeated delays signal unreliability to the team lead.Reassignment of tasks: If a task isn’t finished on time — especially if the annotator stops responding — it may be reassigned, and payment goes to the new person.Working with Deadlines EffectivelyClarify deadlines for each task. If unclear, ask directly.Estimate realistic completion time based on volume and difficulty. Remember: speed is slower at the beginning of a new project.Ask the team lead about expected speed (e.g., 100 objects per hour or 50 images per day) — this helps you pace yourself and adjust if needed.If a task becomes too large or a problem occurs — inform the team lead early. This can prevent reassignment and loss of payment.Track your progress. Don’t postpone work. Break it into parts and work steadily to avoid last-minute rushes.Punctuality is part of professional discipline. An annotator who can manage time properly not only helps the team run smoothly but also earns trust from team leads. Punctuality is as important as accuracy — and often what sets a professional apart from a beginner.Communication with the Team Lead and ValidatorsAn annotator isn’t an isolated worker, but part of a team. Smooth project progress depends on clear, regular communication with team leads and validators.Annotation projects are rarely perfect from day one — there are clarifications, edits, and edge cases. The faster and more accurately an annotator communicates, the more valuable they are.What Communication InvolvesClarifying unclear points: If the instructions are confusing — ask the team lead instead of guessing and risking rework.Reporting issues: Interface bugs? Poor image quality? Upload problems? Report them immediately.Flagging edge cases: If an object or case is not covered in the instructions — take a screenshot and ask for guidance. This avoids inconsistent annotation.Responding to feedback: If asked to revise a task, don’t delay. Clarify unclear comments to save time for both sides.Communication ChannelsTeam chats and groups: For general updates, instruction changes, deadlines, and announcements.Direct messages: For personal task-related questions or explanations.System comments (e.g., CVAT): Built-in tools like review mode help anchor feedback directly to annotations.Why It MattersStandardizes work: The more annotators in a team, the more critical consistency becomes. Team leads and validators ensure this.Builds trust: Team leads need to see annotators as careful, responsible, and capable of asking the right questions.Communication isn’t a secondary skill — it’s an essential part of annotation. An annotator who can ask questions, clarify edge cases, align with guidelines, and respond to feedback is a valuable team member.Self-validationBefore submitting a task for review, an annotator can and should self-check their work. This is part of a professional approach.What to Check:All required objects are annotated; no image is left out.Classes are correctly assigned (e.g., a pedestrian isn’t marked as a cyclist).All instructions are followed, and the right tools were used.Shapes and boundaries match the object’s geometry.No duplicates exist — each object should be annotated only once, unless multiple tools are required.Self-validation isn’t a waste of time — it’s an investment in quality. It reduces rework and increases the chance of promotion to roles like validator or team lead.The Annotator’s ToolkitAnnotation is a task that requires precision. And it’s not only about what you annotate, but also what tools you use to do it.Computer MouseOne of the most essential tools. A good mouse should be:Accurate – it should not “jump” on the screen during small or slow movements.Responsive – it should work without delay.Predictable – cursor movement should match hand movement.Recommendations:Use a gaming mouse (entry- or mid-level) — these are usually accurate and have quality sensors.Use moderate sensitivity for annotation — the cursor should move smoothly and precisely.Avoid cheap office models that stutter or lose tracking.If possible, choose a wired or high-quality wireless mouse.Mouse PadEven the best mouse can perform poorly without the right surface. Some mice (especially optical ones) don’t work well on glossy, glass, or overly smooth tables — the cursor may drift, freeze, or jump.A good mouse pad provides stable and predictable movement and helps prevent jittery annotations or accidental jerks. A mouse pad isn’t just an accessory — it’s an essential tool for stable and precise annotation work.MonitorTo annotate precisely, you must see details clearly.The higher the resolution, the better.IPS panel is ideal — it offers accurate colors and wide viewing angles.Avoid annotating on a small laptop screen — important details can be missed.WorkspaceComfortable hand and mouse positioning.Stable surface.Minimal distractions.Why the Annotator Determines the Project’s SuccessThe annotator is a key part of the entire data processing pipeline. It’s at the annotation stage that the quality of the training material is established — material that will later be used to train machine learning, computer vision, and other AI systems.If annotation is done inaccurately, inconsistently, or with errors, even the most advanced AI model won’t be able to train correctly. As a result, not only does the final product suffer, but so does the entire team — from validators to engineers and clients.The Critical Role of the AnnotatorThey shape the “learning reality” — each annotation becomes part of the dataset the AI will “learn” from. An annotator’s mistake becomes a repeated error in the model.They ensure accuracy and structure — the annotator decides where an object begins and ends, which class it belongs to, and how it should be labeled. These decisions directly influence the model’s behavior in real-world scenarios.They set quality standards — the higher the accuracy and consistency of annotations, the less time is needed for validation and rework. This speeds up the team’s workflow and reduces costs.They impact deadlines and the project’s reputation — slow, inconsistent, or error-prone annotation delays the entire process. In large-scale projects, even small delays can lead to missed deadlines and strained client relationships.Why Annotators Are Not “Temporary Workers” but SpecialistsAn annotator is someone who:Understands the project and its goals.Recognizes the value of accuracy and quality.Can follow guidelines and standards.Engages in ongoing communication with the team.Is able to analyze, clarify, and adapt.The more responsibility an annotator takes on, the more trust they earn — and the more complex tasks they can be assigned, from simple bounding boxes to multi-layered semantic segmentation.Project success begins with the annotator. They are the ones who lay the foundation for all further work.A responsible, attentive, and well-trained annotator is an asset — not just a worker. The effectiveness of the model, client satisfaction, and the overall success of the project all depend on them.Annotator — A Profession Shaping the Future of AIBeing an annotator is not a temporary role or a mechanical task. It’s a profession that requires attention to detail, responsibility, precision, and a mindset of continuous learning. Annotators create the very foundation on which intelligent systems are trained — from autopilots to medical algorithms.An annotator is a specialist. Their contribution to the project should never be underestimated. Their professionalism directly influences data quality, model effectiveness, deadlines, and the team’s reputation.The annotator is the one who makes AI smart. And that is a big and important job.

Data confidentiality is a principle that ensures information is protected from unauthorized access, use, disclosure, or distribution. This is especially important when working with personal or sensitive data, where mistakes can lead to data leaks and legal consequences.In the context of data annotation, this means:Access on a Need-to-Know BasisAnnotators should only access the data necessary for completing their tasks. They must not share this data with others or use it outside the scope of the project.Limiting DistributionData—especially if it includes human faces, addresses, numbers, or internal documents—must not be copied, published, or used for unrelated purposes.Compliance with Non-Disclosure Agreements (NDAs)Many projects include agreements requiring annotators to keep information confidential. Violating such agreements can lead to serious consequences.Why Data Confidentiality Matters for AnnotatorsLegal Responsibility‍Annotators working with personal or sensitive data may be held accountable for any leaks. If annotation is conducted under a contract or NDA, breaches of confidentiality may result in fines, termination, and legal action from the client.Professional Reputation‍A data leak caused by an annotator can seriously damage their reputation as a professional:the client may choose not to work with them again,the information may be shared with other employers or platforms,access to new projects—especially confidential or high-paying ones—may be restricted.Violation of Professional Ethics‍Data annotation requires precision and responsibility. Leaks undermine trust from:the client who provided the materials,the team that expects mutual rule-following,the individuals depicted in the data (e.g., photos or videos of people).Impact on Career GrowthResponsible handling of data is a key indicator of a mature professional. Annotators who strictly follow confidentiality rules:gain access to more complex and restricted projects,may be promoted to validator or team lead roles,are more likely to receive referrals.Maintaining confidentiality is not just a formality. It’s the foundation of trust in the annotation profession, a way to protect yourself and your team from complications, and a chance to demonstrate competence and reliability.What Kind of Data Is Considered Confidential in Annotation?Confidential data refers to information that must not be disclosed, copied, used outside the project, or shared with third parties. Such data may involve both personal and commercial secrecy.Main types of confidential data in annotation projects:Personal Data Information that can directly or indirectly identify a person, including:faces in photos or videos,passport information, phone numbers, addresses,name tags, badges, license plates,voice recordings.Medical and Biometric Information Projects involving medical images or biometric data (e.g., facial scans, iris scans, fingerprints) require special caution:such data is protected by law (e.g., HIPAA, GDPR),even anonymized images can be sensitive.Internal Company Data Information related to internal processes, technologies, or logistics:footage from factory cameras,software interfaces, databases, diagrams,blueprints, device screens, workplace documents.Data Related to Secured Facilities or Restricted Areasbuilding plans, secured zones, security systems,location of surveillance cameras or guards,license plates of service vehicles.Commercial and Intellectual Propertyproduct prototypes, logos, project names,documents marked as “confidential” or “internal use only,”elements not yet released to the public (e.g., products in development).Important to Remember: Even if the data "doesn’t seem important," it might still be confidential. If you’re unsure, it’s always best to consult your team lead or project manager for guidance on handling specific types of data.How Annotators Can Ensure Data ConfidentialityWork Only in a Trusted EnvironmentUse only official accounts and devices approved for the project.Do not download or store data on personal phones, laptops, or USB drives unless explicitly allowed.Avoid working in public places (cafés, co-working spaces), especially with open Wi-Fi networks.Do Not Take Screenshots or Record the Screen for Personal UseAnnotators are strictly prohibited from saving images, videos, or parts of the project interface on their devices for personal purposes — this may lead to a data breach. Screenshots are allowed only with the team lead's permission and strictly for work-related needs, such as asking a question or reporting an error. Always use official project communication channels for this.Do Not Discuss Data Outside the Work EnvironmentNever share images or links in personal chats, Telegram, Discord, etc.Do not discuss project content with friends, family, or annotators not involved in your project.Follow the NDA (Non-Disclosure Agreement)If you signed an NDA, follow its terms precisely.Even after the project ends, you are obligated to keep the information confidential.Use Corporate Tools OnlyWork within official platforms, such as CVAT.You are granted access to the task only through specific platforms. Downloading data and annotating it outside the system is strictly forbidden. All work must be done within the official project environment.All discussions should take place only in work-related chats — Telegram groups, Slack channels, or other official communication tools.Do Not Use Project Materials in Your PortfolioDo not post data or annotation screenshots in your resume, social media, or public profiles.If you need to showcase your experience, use anonymized and pre-approved examples.What to Do If You Suspect a Data LeakReport It Immediately to the Responsible PersonIf you notice something suspicious — such as unauthorized access, unknown copies of data, or unethical behavior from a team member — report it immediately to your team lead or project manager. The sooner the leadership is informed, the more effective and timely the response can be.Do Not Try to Fix the Situation on Your OwnAttempting to "fix" or delete files yourself can make things worse. This may result in:deletion of important evidence,disruption of the investigation process,escalation of the data breach. Wait for instructions from your manager or the security team.Pause Work with Suspicious Data or DevicesIf you suspect that a file, platform, or device might be compromised, stop working with it until you receive further clarification. This can help prevent the spread of confidential information.Maintain Confidentiality While Discussing the IncidentDo not share details about the suspected breach outside of the small group responsible for project security. Avoid discussing it in public chats, social media, or messaging apps. This helps prevent panic, rumors, and unauthorized data exposure.Cooperate with the InvestigationIf an internal investigation is initiated, fully cooperate and provide any necessary information — such as access logs, details of your work with the data, etc. This will help identify the source of the problem and take corrective action faster.Why Is Proper Response Important?Quick action helps minimize damage and prevents similar incidents in the future.Following protocol maintains client trust and the team’s reputation.Correct behavior ensures compliance with laws and the company's internal policies.Legal Aspects of Data Confidentiality for Annotators: USA and EuropeModern data annotation requires strict compliance with privacy regulations, especially when handling personal or sensitive information. Annotators working with projects from the US or Europe must understand the core legal frameworks governing data protection to avoid serious legal consequences and ensure data security.Europe — General Data Protection Regulation (GDPR)What is GDPR?‍The General Data Protection Regulation (GDPR) is the main data protection law in the European Union. Enforced since 2018, it is one of the strictest privacy laws in the world.Examples of data considered confidential under GDPR:Photos and videos with recognizable faces — these are biometric data.IP addresses and cookies, if they can be used to identify a user.Voice recordings — even short clips may qualify as personal data.Medical records — including scans, diagnoses, and prescriptions.Automated user profiles — such as classifications based on interests or behavior.What annotators need to know:Even if a person in a video is not named, if they are recognizable, it qualifies as personal data.GDPR requires anonymization or pseudonymization of data before processing.Annotators may not use data outside the project — not even for portfolios or training purposes.United States — Sector-Specific and State-Level RegulationsUnlike the EU, the US does not have a single, unified law like GDPR. Instead, it follows industry-specific and state-level regulations.HIPAA (Health Insurance Portability and Accountability Act)Who is covered?‍Medical institutions and anyone handling health-related data.What is considered confidential?Patient namesMedical images (e.g., MRI, X-rays)Medical record numbersPhotos where a patient can be identifiedFor annotators:‍If you work with medical imagery from the US, you must hide or ignore any personal identifiers.COPPA (Children’s Online Privacy Protection Act)Who is covered?‍Anyone collecting data from children under 13 years old.Examples of confidential data:A child's face in a photo or videoTheir voice, name, or locationAny account activity related to a childFor annotators:‍Be especially careful when working with children's data. Such data almost always require parental consent and enhanced protection.California Consumer Privacy Act (CCPA) and CPRAWhat is protected?‍Personal data of all California residents.Examples:Names, email addressesGeolocationOnline purchase and behavior dataBiometric informationUser rights under CCPA/CPRA:Know what data is collectedOpt out of data sellingRequest deletion of their dataFor annotators:‍If your project involves users from California, it is critical not to store, share, or replicate data unnecessarily.Key Takeaways for AnnotatorsIt doesn’t matter where you are located — what matters is where the data originates.If you process data from EU citizens, you are required to follow GDPR.If you work with US-based data, make sure you understand the relevant sectoral and state-level regulations.Security and confidentiality are not optional — they are core professional responsibilities.Data Breaches: Real Cases and ConsequencesHeart of England NHS Foundation Trust Employee‍An employee accessed medical records of 14 individuals, including family and friends, without authorization. She was found guilty of violating the UK Data Protection Act and fined £1,000, plus court costs.RAC Employee‍An employee unlawfully collected and passed client data to third parties, which led to unwanted phone calls. She received an 8-month suspended prison sentence and was ordered to pay £25,000.Carlos Lopez & Associates Employee‍An employee accidentally emailed a spreadsheet with personal data of 130 current and former employees, including Social Security numbers and addresses, to 65 colleagues. Despite no proven misuse, three victims filed a class-action lawsuit claiming risk of identity theft. The court ruled that even potential harm is grounds for legal action.WM Morrison Supermarkets Data Leak‍In 2014, an internal audit employee working remotely copied and leaked personal data of nearly 100,000 employees on a public file-sharing platform. He was sentenced to 8 years in prison. The company also faced a collective lawsuit from affected employees.These cases underscore the importance of following security protocols and raising awareness among team members. Unauthorized access — even accidental — can lead to severe consequences.ConclusionIn today’s world, data confidentiality is not just a legal requirement but a fundamental part of professional ethics and safe information handling. For annotators, following confidentiality rules is essential — it protects not only sensitive and commercial data but also your own professional reputation and that of your team and organization. Understanding the types of confidential data, working with discipline, and knowing key legal frameworks such as GDPR and HIPAA help minimize risks of data breaches and legal liability. Real-life cases show that even unintentional mistakes can result in serious sanctions and damage. Paying careful attention to confidentiality is the foundation of a successful, secure annotation career — and the key to earning client trust and long-term professional growth.‍

CVAT (Computer Vision Annotation Tool) is a modern and flexible web-based tool for annotating images and videos used in computer vision tasks. CVAT provides manual, automatic, and semi-automatic annotation capabilities, allowing users to define object boundaries and classes, which are essential for training and testing machine learning models.In CVAT, the annotation workflow is organized in a hierarchical structure: Project → Task → Job. At the top level are projects, within which individual tasks are created. Each task can be divided into several jobs to facilitate distribution among annotators. Annotation is always performed at the job level; jobs are an integral part of a task and cannot be deleted or exist independently.After logging into the CVAT interface, the Tasks page opens by default, displaying all tasks created by the user or assigned to them. From this page, users can also navigate to other sections, including Projects and Jobs.In addition to the main sections, the interface also includes several additional tabs:Cloud Storages — used to connect external cloud storage services (More details: Cloud storages page | CVAT);Requests — allows tracking the status of operations such as task creation, dataset export, or import (More details: Requests page | CVAT);Models — contains pre-installed models for automatic or semi-automatic annotation, and also allows uploading custom models (More details: Models | CVAT).In this material, we will focus primarily on the key sections: Projects, Tasks, and Jobs, as well as the annotation interface, since these form the core of an annotator's daily work. We will now take a closer look at each of them.ProjectsIn the CVAT system, the workflow is organized through projects. Projects play a key role in organizing, standardizing, and scaling the annotation process, especially when teams of annotators, validators, and team leads are involved. (More details: Projects page | CVAT)Centralized Task Management A project in CVAT serves as a container that groups together all tasks related to a specific topic, data type, or research objective. This simplifies workflow navigation, especially when managing a large number of tasks and team members.Unified Annotation SchemaOne of the main advantages of projects is the ability to define a unified label schema that is automatically applied to all tasks within the project. This ensures annotation consistency and helps prevent errors caused by inconsistent terminology or accidental use of different classes for the same object.Team Collaboration ManagementProjects make it easy to organize collaborative work: annotators, validators, and team leads can be assigned, with roles and access rights distributed accordingly. This makes the system especially useful for companies and research groups working on large datasets.TasksTasks occupy a central place in the CVAT structure and serve as the primary unit of data processing. Each task groups together the uploaded images or videos. (More details: Tasks page | CVAT)Content and Structure of TasksA task contains a set of media files to be annotated and inherits the label schema from the project it belongs to. If a task is created outside of a project, the label schema is defined individually. Thus, a task specifies what exactly needs to be annotated and according to which rules. (More details: Task details | CVAT)Subdivision into Jobs and Progress TrackingAfter creation, a task is divided into one or more jobs. This allows the workload to be distributed among multiple annotators or the annotation to be performed in parts. The division is based on the number of frames or images, according to manually set parameters during task creation. The task interface allows tracking of progress status, annotator activity, and annotation quality (when Ground Truth or HoneyPots are available).JobThe Jobs page displays all available jobs and also serves as a personalized list of all jobs assigned to a specific user. It shows tiles for each job, with each tile corresponding to a separate segment of a task designated for annotation.This section gathers all active jobs assigned to the annotator, which may belong to different tasks and even different projects. This is convenient because the annotator doesn’t need to navigate through individual tasks on the Tasks page to find their active assignments — everything they need is collected in one place. (More details: Jobs page | CVAT) Navigating to Annotation Clicking on any job tile opens the annotation interface. This is the main workspace where all annotation work is performed.Convenient Work Distribution Jobs allow the client or task coordinator to evenly distribute the workload among multiple annotators. Each annotator receives only the jobs assigned to them and sees these jobs on their personal page. This reduces the risk of overlap and errors in task assignment.‍Annotation InterfaceThe annotation interface is the central workspace in CVAT where data annotation is performed. You can access it in two ways: by selecting the desired job on the task page or by opening it directly from the Jobs page, which displays all the jobs assigned to the user.The annotation interface combines various panels and tools, organized to enable the annotator to work sequentially and efficiently — from reviewing requirements to creating, editing, and verifying objects.It is important not only to know where the buttons are located but also to understand how the interface is designed from a practical perspective. Familiarity with its features, including settings, keyboard shortcuts, and filters, helps speed up the annotation process, avoid mistakes, and focus on the task content rather than technical details. (More details: CVAT Annotation Interface)‍1. SettingsCVAT offers a wide range of options with a variety of settings. To open the settings panel, simply click on your username in the top right corner and select Settings from the dropdown menu, or use the keyboard shortcut F2. (More details: Settings | CVAT)In the settings, you can adjust video playback parameters: change the player speed, set the frame step, enable or disable autosave, and configure the display of information about annotated objects. Additional options are also available here to activate or deactivate various CVAT features, which is especially convenient when working with different types of tasks.An additional section, Appearance, located on the objects sidebar, allows you to customize the visual style of annotations, including colors, transparency, and other visual parameters of the objects. (More details: Objects sidebar | CVAT #Appearance)There is also an option to adjust the brightness and contrast of the image. These settings are located at the bottom of the workspace and can be accessed by clicking the arrow. The option to enable a grid is also available here. (More details: Workspace | CVAT #Image settings)2. Reviewing Task Requirements and Navigation InterfaceBefore starting annotation, it is important to carefully review the instructions and task parameters. This helps the annotator understand which objects need to be annotated, which classes and attributes to use, as well as the quality and formatting requirements for the annotations.The top panel of the interface includes several key elements:The Guide button opens a detailed instruction for the task. It describes which objects need to be annotated, their structure, the available classes, and special considerations to keep in mind during annotation.The Info button displays information about the task: total number of frames, start and end frames, the user assigned to the job, as well as statistics on annotated objects by class. This allows for a quick assessment of progress and workload.The Filters button opens the object filtering menu. This feature allows you to filter displayed annotations by label, shape, or other criteria. It is especially useful during task validation but also helps during annotation — for example, for self-checking or navigating existing objects.The annotation mode is available via a dropdown menu next to these buttons. This menu allows switching between different modes depending on the annotator’s current tasks: Standard mode — the default mode where objects are created and edited directly on frames. It is suitable for most tasks and serves as the main working mode. In this mode, the annotator can freely switch between tools, create new objects, edit existing ones, and review annotations.Attribute annotation mode — used for bulk or step-by-step filling of object attributes. Useful when the focus is on metadata rather than geometry.Single shape mode — restricts the user to creating one object of a specified class per frame, after which the frame automatically switches to the next.Tag annotation mode — allows adding tags to images and is used for more convenient and efficient annotation of image classification tasks.Review mode — designed for annotation validation. This mode is convenient for checking already annotated objects, identifying errors, and leaving comments.Each mode is designed for specific scenarios, and choosing the appropriate mode at the right time helps significantly simplify and speed up the workflow. (More details: Navigation bar & Menu | CVAT #job information and annotation mode switcher)Understanding these interface elements and using them correctly is key to ensuring that the annotation meets the client’s expectations, is of high quality, and complies with all project requirements.‍3. Selecting the Tool and LabelAfter reviewing the instructions, the annotator selects an annotation tool — the left panel offers tools such as rectangle, polygon, polyline, points, and others. Semi-automatic and automatic annotation modes can also be used by selecting AI Tools.On the tools sidebar, the annotator chooses the desired shape (rectangle, polygon, or another), selects the label, and then selects the annotation mode: Shape or Track.In Shape mode, we usually annotate individual images.In Track mode, we work with video clips where it is necessary to track objects that appear and move throughout the video.More details:Controls sidebar | CVAT #Shapes Shape mode (basics) | CVAT Shape mode (advanced) | CVAT Track mode (basics) | CVATTrack mode (advanced) | CVAT ‍4. Annotation in the Workspace and Using ShortcutsThe central part of the screen is the workspace, where the annotator views images or video frames and performs the main annotation work. This is where annotations are created. (More details: CVAT Workspace) Depending on the data type, different approaches are used: annotating each image individually, interpolation, or tracking.If needed, images can be zoomed, moved, or rotated — navigation is available both via mouse and keyboard shortcuts.More details: Shortcuts | CVAT It is also worth highlighting the shortcut key settings. The user can remap almost any key combination, making annotation more convenient and faster. To view the current shortcuts, simply press F1 — a window will open displaying a list of all available shortcuts.Shortcuts can be modified in the Settings under the Shortcuts tab.Using shortcuts significantly speeds up the annotation process, as you don’t waste extra time moving the mouse pointer to specific buttons. Learning and using shortcuts is especially important for improving productivity.5. Task NavigationThe annotation interface includes a player for navigating through the task. It allows you to view video, move frame-by-frame, jump forward or backward by a specified number of frames, and navigate to the beginning or end of the task. The player bar helps track task progress and quickly jump to an approximate position. By knowing the frame number, you can move directly to it by entering the number or by specifying the file name. All these features enable efficient navigation through the task during annotation or validation.More details: Navigation bar & Menu | CVAT #Navigation controls 6. Working with the Objects sidebarAll annotated objects are displayed on the right side of the screen — in the Objects sidebar. Here you can quickly see the list of all annotations present on the current frame, change their labels, and set or refine attributes.Various properties applicable to the objects are available here. Objects can be sorted using different methods.There are additional tabs: Labels and Issues.In the Labels tab, we can see all the labels available in the task, while the Issues tab shows problems found in the task. Issues are added by the validator during task review.More details: Objects sidebar | CVATManual QA and Review | CVAT ‍7. Progress Monitoring and Task CompletionUsing the information from the player bar, you can determine when you have reached the end of the task. When finishing the annotation, the annotator should change the task status accordingly in the menu, then save and close the task.The menu also contains other useful functions that assist with annotation or annotation management, such as Remove annotation and Run actions. However, these should be used carefully to avoid accidentally deleting or modifying annotations that do not require changes.More details: Navigation bar & Menu | CVAT #Menu ‍ConclusionCVAT is a powerful and flexible tool for annotating images and videos, supporting a wide range of workflows — from manual annotation to automated processes using artificial intelligence models. Its structured approach to organizing the workflow through projects, tasks, and jobs allows for efficient management of large volumes of data and the distribution of work among team members.The CVAT interface is designed to provide annotators with convenient and functional tools for fast and high-quality annotation. The ability to customize keyboard shortcuts, different annotation modes, as well as easy navigation and progress tracking significantly enhance productivity and work quality.Thus, a deep understanding of CVAT’s structure and skillful use of its features is the key to successfully completing annotation tasks, ultimately contributing to the creation of accurate and reliable datasets for training modern computer vision models.

Getting Started with CVAT Online (Part 1)

Getting Started with CVAT Online (Part 2)

Hello, and welcome! Today, we’re diving into one of the most essential tools in data annotation for computer vision—bounding boxes. Whether it’s detecting objects in images, classifying them, or tracking their movement in videos, bounding boxes form the foundation of many machine learning models. But what makes them so effective, and why are they still widely used despite more advanced techniques? Let’s find out.What is the Bounding Box?A bounding box is a fundamental data annotation tool used in computer vision tasks. It is a rectangular frame that encompasses an object in an image or video stream, defining its location and size. A bounding box is created using two or four coordinates, making the annotation process quick and convenient. This method is widely applied in training machine learning models for object detection, classification, and tracking tasks. Despite its simplicity, a bounding box provides sufficient accuracy for most tasks.Why is the Bounding Box effective?Bounding box is popular and effective for the following reasons:Speed and simplicity. You can build a bounding box in a few seconds using only 2 points: the upper left and lower right corners of the frame. This makes the annotation process faster and adaptable to large amounts of data, reducing the cost of partitioning.Suitable for most objects. Many objects in images can be framed by a rectangle with minimal loss of context, allowing models to adequately recognize and interpret objects. For example, cars, people, animals, and objects can be easily highlighted with a rectangular frame, allowing the model to be accurately trained.Reduced computational complexity. Bounding box requires less data to describe an object than, for example, contour or mask partitioning. This reduces the load on the model training process and allows faster processing of datasets while maintaining high accuracy in object classification and detection.Bounding boxes are great for tasks where it is important to simply select an object without accurately outlining its shape, which is especially useful for training models under time and resource constraints.Using the Bounding BoxBounding box is actively used for many tasks, optimizing its application in complex cases. Bounding box is most often used for:Object detection and classification. Bounding box helps to accurately label an object and save its characteristics such as location and size. This allows you to train models to find and classify objects in images and videos, which is important for automated security, object recognition and navigation systems.Motion analysis and tracking. In the task of tracking objects in video streams, bounding box is used to determine the coordinates of the object in each frame. This allows models to identify objects and track their motion. It is important to minimize the inclusion of background so that the model can extract important characteristics of the object, such as its contours and texture, and use this data for subsequent identification.Contemporary Challenges and Alternatives to Bounding BoxBounding box is a universal tool, but it is not without its limitations. This is primarily true for objects with irregular shapes ( for example, animals or trees), where a simple rectangular box may capture too much unnecessary background. In such cases, models may require more detailed annotation, such as mask segmentation or segmentation using polygons, to improve accuracy. Using a mask or polygon to precisely draw the outline of an object allows you to capture only the object and get rid of unnecessary background.Black - bounding boxRed - captured background by bounding boxGreen - mask segmentationHowever, for massive detection tasks, bounding box remains the standard because it offers a balance between simplicity and sufficient accuracy. It is convenient for data-intensive annotations, such as in observation systems or for training object recognition models.Examples of successful usageBounding box is widely used in large datasets for computer vision tasks, especially for object detection, classification and tracking. These datasets are important for model training because they provide a variety of annotated images that allow models to learn object recognition efficiently. Here are some of the best known datasets where bounding box is used:‍COCO (Common Objects in Context)‍The COCO dataset is one of the most popular in computer vision. It includes annotations for over 300,000 images and supports several types of partitioning, including bounding box and object segmentation. COCO includes objects from more than 80 categories, making it a versatile tool for training models that solve detection, segmentation, and even image description problems. COCO provides annotations for objects in the natural environment, and the bounding box allows models to accurately locate objects. This dataset is used for many competitions and scientific research, as it includes complex scenes with overlapping and interacting objects.‍Pascal VOC (Visual Object Classes)‍Pascal VOC was one of the first datasets to make extensive use of bounding boxes for annotation. It contains images with objects of 20 classes and offers annotation in the form of bounding boxes and masks. Pascal VOC has played a significant role in the development of object recognition algorithms as it includes scenes of varying complexity, from simple to highly loaded, allowing models to learn to select objects under near real-world conditions. This dataset is still used as a basis for comparing models and testing object detection methods, although it is smaller in volume compared to COCO.‍ImageNet‍ImageNet is one of the largest and best known datasets, covering thousands of object categories. In addition to classification, ImageNet provides bounding boxes for most categories, which allows it to be used in object detection and localization tasks. The ImageNet dataset is used as a standard for training and testing models, and while the primary focus is on classification, the bounding box allows models to be trained to accurately localize objects. This approach is used for tasks where it is important to both recognize an object and accurately indicate its location in the image.Open Images Dataset‍Open Images is another large dataset from Google, containing more than 9 million images and hundreds of thousands of annotations in bounding box format. The dataset includes over 600 object classes and provides annotations not only in bounding box format, but also as masks for segmentation and relationships between objects (e.g., interactions between people and objects). This makes Open Images useful for training models that need not only to identify an object, but also to understand its environment and interaction with other elements in the scene.In summary, a bounding box is a simple but effective tool that allows you to quickly and accurately label objects in images and videos. Its application is an important part of the process of creating high-quality datasets. Given the current and future challenges in AI and computer vision, bounding box is likely to remain a key element in data annotation, striking a balance between simplicity, accuracy and speed of annotation.

Bounding Box Annotation in CVAT (Overview)

Bounding Box Annotation in CVAT (Practical Task)

Polygon AnnotationPolygon is an annotation tool used to create a closed polygon that outlines the contours of an object. Each vertex of the polygon is created manually, enabling precise delineation of the object’s shape, including its complex and irregular contours. A polygon consists of a sequence of points connected by lines, with the last point closing the contour.Polygon allows you to accurately outline an object, excluding the background.Polyline AnnotationPolyline is a tool for creating an open line, used to annotate objects where the path or structure is important, but a closed contour is not required. For instance, polylines are used for annotating roads, route boundaries, or structural lines such as electrical wires or rivers. Each polyline point is added manually, forming a sequential path.Polyline is used for linear objects where path is important, not area.Tasks for Polygon and PolylineDifferent tasks - different tools: Polygon for shape, Polyline for structure.Polygon‍The Polygon tool was designed for tasks requiring high precision in describing object boundaries. Rectangular bounding boxes are not always suitable, especially for objects with complex shapes such as animals, plants, or items with multiple angles. The polygon tool allows you to:Eliminate unnecessary background.Capture only the essential details of the object.Improve model accuracy for tasks where pixel-level precision is crucial, such as object segmentation.Polyline‍The Polyline tool is used for tasks requiring the annotation of paths, routes, or lines. This tool is indispensable for describing structural objects that are not solid figures. Examples include:Tracing roads or road markings in navigation systems.Defining boundaries of fields or territories.Annotating power lines. Polyline is essential where the structure matters, not the area of the object.Efficiency of Polygon and PolylinePolygonAdvantages:High precision: Enables detailed descriptions of complex object shapes.Minimizes background: Excludes unnecessary pixels, improving model training.Flexibility: Suitable for any objects, regardless of their shape or size.Limitations:Time-consuming: Annotating complex objects takes more time than using bounding boxes.Challenging for large datasets: The process becomes labor-intensive for large datasets.PolylineAdvantages:Ideal for structural objects: Perfect for linear objects such as roads, routes, or building outlines.Ease of use: Drawing lines takes less time than detailed polygon annotations.Limitations:Cannot describe areas: Unsuitable for tasks requiring the complete shape of an object.Limited application: Used only for objects with linear structures.Polygon Example: Annotating a Tree Leaf with Jagged EdgesIn plant classification tasks, annotators often need to mark individual leaves that have sharp, uneven edges. Using a bounding box would include too much irrelevant background and lose key features like shape and edge curvature.Why Polygon works: It allows precise capture of the object’s true outline, improving model accuracy where pixel-level details matter.Trade-off: Time-consuming to annotate, especially across large datasets.Polyline Example: Mapping a Pedestrian Path on Aerial Images.For city navigation datasets, annotators draw paths along walkways, roads, or trail networks. The goal is to mark the direction and course of travel, not the entire area.Why Polyline works: It efficiently traces the route with minimal effort, ideal for large-scale projects.Limitation: Cannot describe area or full shape, suitable only for linear features.‍Examples of Use in Popular DatasetsPolygonCOCO (Common Objects in Context)Use of Polygons: COCO includes polygon annotations for object segmentation tasks. This enables precise outlining of object boundaries, crucial for instance segmentation.Type of Annotations: Instance Segmentation (polygons for each object).Examples of Objects: People, cars, animals, household items.Number of Images: Approximately 330,000 images, with over 200,000 annotated with polygons.Applications: Development of models for Mask R-CNN and other segmentation methods.Accurate annotations of complex objects that overlap.Feature: Polygons are annotated with high precision, including intricate shapes and objects with numerous details.CityscapesUse of Polygons: Cityscapes includes annotations using polygons for semantic and instance segmentation in urban environments.Type of Annotations: Instance Segmentation and Semantic Segmentation (polygons for objects and regions).Examples of Objects: Cars, pedestrians, buildings, roads, trees.Number of Images: 5,000 images with fine annotations.20,000 images with coarse annotations.Applications: Autonomous driving, requiring precise identification of urban infrastructure.Urban scene analysis with accurate delineation of object boundaries.Feature: Highly detailed polygons, especially in complex scenes with numerous objects, such as intersections.Pascal VOC (Visual Object Classes)Use of Polygons: Pascal VOC provides polygons for object segmentation tasks. Though smaller than COCO or Cityscapes, it remains valuable for basic tasks.Type of Annotations: Semantic Segmentation (polygons for objects).Examples of Objects: Animals, vehicles, furniture, people.Number of Images: 11,530 images.Applications: Training models for basic object segmentation.Annotating objects in small datasets.Feature: Simplicity and smaller data volume allow for quick testing of models on polygonal annotations.PolylineCityscapesPolylines in this dataset are used for annotating road markings, sidewalks, lanes, and object boundaries.Object Classes: Roads, sidewalks, markings, buildings, boundaries of road elements.Number of Images: 5,000 images with fine annotations.20,000 images with coarse annotations.Features: Suitable for autonomous driving tasks. Polylines are annotated with high precision, allowing models to effectively analyze urban scenes.BDD100K (Berkeley DeepDrive)Polylines are used to annotate road markings, lanes, and road boundaries.Object Classes: Road markings, lanes, boundaries of objects.Number of Images: Over 100,000 images.Features: A large dataset with diverse road scenes, including various weather conditions and times of day.OpenLanePolylines are used to annotate lanes, motion trajectories, and boundaries.Object Classes: Lanes, road markings, boundaries of objects.Number of Images: Approximately 200,000 images.Features: Specialized for autonomous driving tasks with detailed annotations of lanes.Conclusion‍Polygon and Polyline are key annotation tools, enabling tasks where precision and structure are critical. Polygon is ideal for objects with complex shapes that require detailed segmentation, while Polyline is effective for linear structures such as roads or boundaries. These tools are actively used in modern datasets like COCO and Mapillary Vistas, underscoring their importance for training machine learning models. When choosing an annotation tool, it is essential to consider the specific task to ensure maximum accuracy and efficiency.‍

Polygons & Polylines in CVAT (Overview)

Polygons & Polylines in CVAT (Practical Task)

One of the most frequently used annotation tools in CVAT is the Brush tool. This tool combines the capabilities of brush and polygon annotations, providing flexibility in creating masks for objects of any complexity.What is the Brush Tool in CVAT?The Brush tool in CVAT allows you to draw masks manually using the cursor as a brush or create them based on polygons. When using polygons, we mark the outer contours of the object, and this area is then transformed into a mask. This makes the tool especially useful for tasks that require high precision and detailing, including complex or irregular object shapes.Brush Tool (Mask) vs. Polygons: What’s the Difference?Masks and polygons in CVAT have different approaches to annotating objects:Masks allow you to highlight both the outer contours and internal areas of objects. Using the Brush tool, complex shapes can be annotated, including holes and details inside objects.Polygons, on the other hand, are limited to annotating only the outer contour of the object. CVAT does not allow cutting internal areas or creating holes within polygons.This key difference makes masks more versatile in tasks that require detailed work.Advantages and Disadvantages of the Brush ToolAdvantages of the Brush Tool:Flexibility: Allows the creation of masks of any shape, including complex contours and internal holes.High precision and detailing: The Brush tool lets you work with minute details, including hair and other small elements.Disadvantages of the Brush Tool:Labor-intensive: Working with the Brush tool can take a long time, especially for large objects or complex scenes.Skill requirement: Annotators need more experience and attention to detail to work effectively with the Brush tool compared to using a simple polygon.Editing difficulties: Correcting masks can be a labor-intensive process, especially when significant revisions are needed.Main Tasks of the Brush Tool in CVAT‍Masks are a versatile tool for tasks that require high accuracy and detailed object selection. Here are the main examples where masks are applied:Object Segmentation in ImagesMasks are widely used in instance segmentation and semantic segmentation tasks for pixel-level object selection. This is crucial in computer vision for identifying complex shapes, such as people, animals, plants.Background RemovalIn urban datasets, masks are often used to isolate vehicles from the background, such as roads, buildings, and pedestrians. For example, when annotating images for autonomous driving models, accurately separating cars, buses, and bicycles from their surroundings ensures the model focuses on the features of the vehicles themselves, improving object detection and recognition performance.Medical SegmentationIn medical imaging, masks are used to highlight organs, tissues, and pathologies (such as tumors) in CT, MRI, and X-ray images, aiding in diagnosis and research.Examples of Mask Use‍Popular datasets for masks and polygons may overlap, but their use depends on the type of tasks the data is intended for. The primary difference lies in the annotation format and the level of detail required for a specific case.COCO (Common Objects in Context)Supports instance segmentation, providing masks for objects. Masks are important for tasks that require high-precision object segmentation.Example use: training Mask R-CNN.CityscapesContains masks for semantic and instance segmentation of urban infrastructure. Provides high-detail scenes, useful for autonomous driving.Example use: segmentation of roads, pedestrians, buildings.Pascal VOCMasks for semantic segmentation, though less complex than those in COCO.Example use: basic object segmentation tasks.Conclusion‍The Brush tool in CVAT is a powerful tool that provides unique capabilities for annotating objects of any complexity. Although working with this tool may require more time and skill, its versatility makes it indispensable in tasks that require detailed annotations and mask work.

Brush (Mask) Tool in CVAT (Overview)

Brush (Mask) Tool in CVAT (Practical Task)

In the process of data annotation for computer vision tasks, various labeling tools are widely used. Among them, points and skeletons hold a special place, as they allow for describing objects with a high degree of accuracy. These methods are actively applied in pose recognition, object tracking, and other tasks.What are points and skeletons in data annotation?Points are a primitive but flexible annotation tool, consisting of individual coordinates in an image. They can be used to label key points of objects, such as eyes, nose, joints, or other characteristic features.Skeleton is a more complex structure, consisting of a set of interconnected points. It is used to annotate objects with a specific structure, such as human or animal figures, where it is important to capture the positions of joints and limbs.In what tasks are these tools used?Pose Recognition Determining body posture for sports applications, VR, and animation.Analyzing movements in medicine and rehabilitation.Biometric Identification Facial recognition considering the positioning of eyes, mouth, and other features.Gesture navigation and control.Autonomous Systems and Robotics Recognizing the position of hands or other body parts for robot control.Optimizing robot interaction with the environment.Advantages and Limitations of ToolsAdvantagesHigh annotation accuracy: Allows marking even the smallest details of objects.Flexibility: Applicable to a wide range of tasks, from medicine to entertainment.Efficiency: Allows annotating complex structures, such as the human body, with minimal data.LimitationsTime-consuming: Manual annotation can be time-intensive.Data quality demands: Errors in annotation can significantly reduce model accuracy.Limited applicability: Points and skeletons cannot fully describe the shape and boundaries of an object. If the task requires recognizing the type of object (e.g., a car, building, or animal), points alone will not provide sufficient information.Examples of Usage in Popular DatasetsCOCO (Common Objects in Context) Contains skeleton annotations for human pose analysis.250,000 images, 17 key points for each annotated person.Used in pose analysis and motion tracking tasks.MPII Human Pose Dataset One of the most well-known datasets for human pose annotation.25,000 images, 40,000 annotated poses.16 key points for each skeleton.Used for training models for motion analysis.PoseTrack Dataset including annotations for image sequences to track poses in motion.135,000 frames with annotated skeletons.15 key points for each skeleton.Applied to analyze sports movements and gestures.Conclusion‍Points and skeletons are powerful data annotation tools, especially in pose recognition and object tracking tasks. Despite their limitations, they remain an integral part of modern computer vision algorithms. With the development of automated annotation tools and improvements in machine learning models, their application will become increasingly precise and efficient.

Points & Skeleton in CVAT (Overview)

Points & Skeleton in CVAT (Practical Task)

1. What Are Attributes in Data Annotation?In CVAT (Computer Vision Annotation Tool), attributes represent properties or characteristics assigned to specific annotated objects in an image or video. They allow for more detailed object descriptions beyond basic labels and provide a deeper understanding of the data.Examples of attributes:Color: "Red," "Blue," "Green"Condition: "Damaged," "Intact"Type: "Sedan," "Hatchback," "SUV"Action: "Walking," "Running," "Standing"Attributes can be of various types, including text fields, numerical values, checkboxes, and dropdown lists, offering flexibility in describing object characteristics.2. In What Tasks Are Attributes Used?Attributes are used in various computer vision tasks where detailed object information is required:2.1. Object DetectionIn object detection tasks, attributes help specify the characteristics of each detected object.Example:Object: CarAttributes: Color: "Red"Type: "Sedan"Condition: "Damaged"2.2. Semantic SegmentationIn segmentation tasks, where each pixel of an image is classified, attributes help refine details of the segmented areas.Example:Region: RoadAttributes: Surface: "Asphalt"Condition: "Wet"2.3. Behavior Analysis and Object TrackingAttribute values can change over time. For example, in tasks related to human or animal behavior analysis, attributes help describe actions and states dynamically.Example:Object: PersonAttributes: Action: "Sitting" → "Standing" → "Walking" → "Running"3. Advantages and Disadvantages of AttributesAdvantages:Detailed descriptions – Attributes allow adding rich information about objects, improving data quality and increasing model accuracy.Flexibility – A variety of attribute types (text, numbers, checkboxes) make it possible to adapt annotations to project-specific requirements.Improved model training – Attributes help models better distinguish and classify objects based on additional characteristics.Disadvantages:Increased annotation complexity – Adding attributes requires extra time and effort from annotators.Potential for errors – Incorrect or inconsistent attribute assignments can reduce data quality and, consequently, model performance.Larger dataset size – Additional attributes increase the volume of annotated data, requiring more storage and processing resources.5. ConclusionAttributes in CVAT play a crucial role in data annotation, allowing for detailed object descriptions and improving the quality of training models.Despite the complexity of annotation and the increased dataset size, attributes enable models to better understand context and make more accurate predictions. Their use is particularly beneficial in projects that require in-depth analysis of visual information.Maintaining consistency in attribute assignment is essential to avoid errors and enhance dataset accuracy.Thus, attributes serve as a powerful tool for improving the quality of annotated data, helping to build more accurate and reliable computer vision systems.

What Are Tags in Data Annotation?Tags are labels used to add additional information to annotated data. Unlike bounding boxes, polygons, or other annotation shapes that define specific regions, tags apply to an entire image or video frame, providing metadata that describes the overall scene rather than individual objects.Tags help structure data, making it easier to filter and process. They can indicate shooting conditions, scene type, or even technical characteristics of an image.What Tasks Use Tags?1. Image ClassificationIn classification tasks, tags serve as primary labels that indicate an image’s category. For example, if a dataset contains images of urban and rural areas, the following tags can be used:CityVillageSuburb2. Video Annotation and Frame ProcessingWhen annotating videos, tags help classify scenes and frames. For example:Moving cameraStatic shotFast object movementThese tags assist algorithms in determining the complexity of video processing and adapting models accordingly.3. Automatic Data FilteringTags can be used to automatically exclude images from the training set if they don’t meet specific criteria (e.g., "too dark image", "blurry photo").Advantages and Disadvantages of TagsAdvantages:Simplified classification – Tags allow quick categorization of images.Flexibility – Custom labels can be added without modifying the data structure.Filtering and search – Tags help quickly locate the necessary images.Support for automated processing – Specific images can be automatically labeled and excluded from training.Disadvantages:Limited detail – Tags provide information about the overall scene but not specific objects.Potential data redundancy – If certain information is duplicated in object attributes, it may lead to unnecessary data complexity.Examples of Tag Usage in Popular DatasetsCOCO (Common Objects in Context)The COCO dataset includes object annotations but also features metadata similar to tags, such as:"Outdoor/Indoor" (outdoor or indoor scene)"Weather conditions" (e.g., rainy, sunny)Open Images DatasetThis dataset uses tags for image classification and to indicate shooting conditions, such as:"Aerial view" (top-down perspective)"Black and white" (grayscale image)ConclusionTags in CVAT play a crucial role in data annotation, especially in classification and video annotation tasks. They help structure datasets, speeding up filtering and data analysis.

Tags & Attributes in CVAT (Overview)

Tags & Attributes in CVAT (Practical Task)

What is a Cuboid?A cuboid (or rectangular parallelepiped) is a three-dimensional geometric shape with six rectangular faces. In computer vision and machine learning, cuboids are often used as volumetric bounding boxes (3D Bounding Boxes) to represent objects in images or point clouds.Applications of CuboidsCuboids are widely used in various tasks requiring three-dimensional data analysis:Computer Vision: Used for annotating objects in 3D space, such as vehicles, pedestrians, and other scene elements.Autonomous Driving: Helps systems recognize and track objects in real time, which is crucial for the safety of self-driving cars.Robotics: Applied in navigation and environment perception, enabling robots to understand the spatial positions of objects.Augmented and Virtual Reality: Used for spatial analysis and interaction with virtual objects.3D Reconstruction and Modeling: Allows for object size estimation and scene analysis using point clouds.Advantages and Disadvantages of CuboidsAdvantages:Simple Representation: Cuboids are easier to compute and process compared to more complex 3D shapes. They can be defined using just a few coordinate points, simplifying their use in machine learning algorithms.Efficient Annotation: Creating cuboids for objects is simpler than detailed shape modeling, significantly speeding up dataset preparation. This is especially important in autonomous driving and other domains that require large-scale annotation.Versatility: They can be used in various tasks, from computer vision to simulations. Cuboids are applicable in both 2D annotation (projections onto a plane) and 3D analysis, making them adaptable to different scenarios.Disadvantages:Low Accuracy for Complex Objects: Cuboids are not well suited for objects with intricate geometry (e.g., humans or animals) since they do not account for shape details. This can lead to detection errors and make object classification more challenging.Issues with Tilted Objects: Cuboids are often constrained to coordinate axes (Axis-Aligned Bounding Boxes), making them inaccurate when annotating objects at an angle. This can result in parts of the cuboid extending beyond the object's boundaries while leaving empty space elsewhere. To address this issue, Oriented Bounding Boxes or more complex 3D models are used, but they require greater computational resources and more sophisticated processing.Challenges with Occluded Objects: When an object is partially covered by another, correctly placing a cuboid can be difficult. This is especially critical in urban environments or on busy roads, where objects frequently obscure one another.Scalability Issues: In large scenes with high object density (e.g., in urban datasets), using cuboids can lead to overlapping annotations and increased errors in recognition algorithms.Popular Datasets with CuboidsCuboids are widely used in the following datasets:KITTI: One of the most well-known datasets for autonomous driving, containing 3D annotations of objects using cuboids. KITTI provides:~15,000 images with annotations~80,256 3D bounding boxes for objectsnuScenes: A dataset for autonomous vehicles, including annotated LIDAR point clouds with 3D bounding boxes. nuScenes contains:1,400,000 camera images390,000 LIDAR scans1,400,000 annotated 3D bounding boxes for 23 object classesWaymo Open Dataset: A large dataset from Waymo for training autonomous driving models with detailed 3D annotations. The Waymo Open Dataset includes:1,950 segments of 20 seconds each (approximately 200,000 frames)12.6 million annotated 3D bounding boxesThese datasets are widely used in research and development of autonomous driving systems and computer vision, providing rich datasets for training and testing models.ConclusionCuboids are a powerful tool in 3D data analysis tasks. They are widely used in autonomous systems, robotics, and computer vision due to their simplicity and efficiency. However, their use comes with limitations, such as the inaccuracy of representing complex objects. Nevertheless, thanks to popular datasets and advancing technology, cuboids remain a key component in 3D analysis.

Cuboids in CVAT (Overview)

Cuboids in CVAT (Practical Task #1)

Cuboids in CVAT (Practical Task #2)

What is an Ellipse Tool?The Ellipse Tool is an annotation tool in CVAT designed for highlighting oval or round-shaped objects using ellipses. Unlike Bounding Box or Polygons, this tool allows for more precise outlining of circular objects without unnecessary background fragments.In what tasks is Ellipse Tool used?The Ellipse Tool is particularly useful in the following scenarios:Annotating round objects: such as car wheels, human or animal eyes, coins, plates, footballs, and other objects with a well-defined round shape.Markup automation in manufacturing: used when analyzing defects on round parts, such as bearings, pipes, and discs, which helps improve quality control and automate the process.Medical image analysis: used in segmenting tumors, cells, and anatomical structures like pupils, skin spots, or the shapes of internal organs in diagnostic images.Aerial photography and satellite analytics: used for annotating objects like building domes, craters, tanks, and round-shaped agricultural land.Advantages and disadvantages of annotation with ellipses Advantages:Accuracy: An ellipse better fits round shapes compared to Bounding Boxes, excluding unnecessary background pixels.Time-saving: In some cases, annotating with ellipses is faster than manually creating polygonal outlines.Better suited for naturally round objects: For biological or mechanical structures, where shape precision is important.Orientation flexibility: The ability to rotate the ellipse at the desired angle allows for better adaptation to the object.Disadvantages:Not suitable for complex shapes: Objects with many protrusions or irregular shapes require annotating with Polygons.Limited support in some neural networks: Most models are trained on Bounding Box or Polygon annotations, making the use of ellipses less standard.Lack of support in popular datasets: Most widely used datasets do not support ellipses as a separate annotation type.Ellipse support in datasets‍Popular datasets do not contain ellipses as a distinct annotation type. They use formats like Bounding Box, Polygons, Keypoints, and Masks, but ellipses are not included. When exporting from CVAT, ellipses are automatically converted into masks to ensure compatibility with annotation formats supported by most datasets and models.Conclusion‍The Ellipse Tool in CVAT is a convenient solution for annontating round objects, helping to avoid unnecessary background and speeding up the annotation process. It is especially useful in tasks related to autonomous driving, medical diagnostics, industrial analysis, and aerial photography. However, it is important to note that most popular datasets do not support ellipses. When exported from CVAT, they are converted into masks, which may require additional transformations to ensure compatibility with existing neural network models. Despite these limitations, using ellipses can significantly improve annotation accuracy in tasks where the round shape of the object is crucial.

Ellipse Tool in CVAT (Overview)

Ellipse Tool in CVAT (Practical Task)

Track Mode in CVAT is a video annotation tool that helps annotate moving objects faster and more conveniently. Instead of manually labeling each frame, annotators can define key points, and the system will automatically fill in the intermediate frames.InterpolationTrack Mode uses interpolation. Interpolation is a method of automatically generating intermediate frames between annotated keyframes. In the context of CVAT, this means that an annotator marks an object in several keyframes, and the system automatically fills in the object's positions in the remaining frames. CVAT supports linear interpolation, where the object moves in a straight line between two key points.Where is Interpolation Used?Interpolation is useful in various computer vision tasks, especially for video annotation. Some key applications include:Autonomous Vehicles – Tracking pedestrians, vehicles, and road signs. Interpolation helps create high-quality datasets for training neural networks that enable vehicles to predict object movements on the road.Surveillance and Security – Tracking people, cars, and other moving objects in real time. This is crucial for detecting suspicious activities, analyzing crowd behavior, and monitoring traffic.Sports Analytics – Automated analysis of athletes’ movements and game objects (balls, pucks), which is used for tactical analysis and team strategy development.Medical Applications – Tracking organ movements, surgical instruments, and analyzing videos from endoscopic and radiological studies.Advantages and Disadvantages of InterpolationAdvantagesTime Savings – Interpolation significantly reduces annotation time since annotators don't have to mark every frame manually.Smooth Tracking – The transition between frames is fluid, making object movement appear natural, especially in long video sequences.Reduced Workload for Annotators – Using interpolation minimizes repetitive tasks, allowing specialists to focus on more complex aspects of annotation.Improved Annotation Consistency – Automatic filling of intermediate frames reduces discrepancies between sequential annotations.Support for Complex Scenes – When combined with manual adjustments, interpolation helps annotate scenes with multiple moving objects, which is especially important for surveillance and autonomous driving.DisadvantagesErrors in Complex Movements – If an object changes trajectory sharply or its speed fluctuates, linear interpolation may result in inaccurate annotations.Issues with Shape Changes – Standard interpolation methods are not well-suited for objects that change shape (e.g., a person raising their hand or a bending object).Need for Validation – Automatic interpolation does not always produce perfect results, so annotators must manually review and adjust annotations, which can take additional time.Popular Datasets with Interpolated ObjectsKITTI Tracking – One of the most popular datasets for automotive computer vision, containing annotated frame sequences with various moving objects.AICity Challenge – Includes videos from urban surveillance cameras, used for traffic analysis and road safety applications.BDD100K – Provides annotated road object videos in various weather and lighting conditions, useful for autonomous driving tasks.TAO (Tracking Any Object) – A dataset designed for tracking various types of objects in videos, supporting annotation interpolation.ConclusionTrack Mode in CVAT is a powerful video annotation tool that speeds up the annotation process through interpolation. However, like any automated method, it requires validation and adjustment. Interpolation-based annotation is widely used in fields such as autonomous driving, surveillance, and sports analytics, making it a valuable tool in computer vision.

Track Mode in CVAT (Overview)

Track Mode in CVAT (Practical Task)

CVAT supports not only traditional manual data annotation, but also tools for automated and semi-automated labeling, which greatly simplify an annotator’s work. To enable this, CVAT offers a dedicated toolkit — AI Tools. This is a set of technologies for semi-automatic and fully automatic annotation of data directly in the editor interface. Its main goal is to make the process of creating datasets for computer vision model training faster and more efficient. With AI Tools, annotators can delegate tasks such as object detection in frames, mask generation, or video object tracking to a model.Main AI Tool TypesAI Tools in CVAT are divided into three key categories:InteractorsInteractors are semi-automatic annotation tools. They work in close collaboration with the annotator, who marks key points of an object (using Positive and Negative Points), while the neural network completes the mask or polygon.Models used:SAM (Segment Anything Model) — a popular tool for creating masks or polygons around an object.Best suited for: Cases where precise annotation of objects is required, but not by doing it completely from scratch.DetectorsDetectors are tools for fully automated annotation. Integrated models automatically locate objects in a frame, define their boundaries (bounding boxes), and assign labels.Models used:YOLO — a popular detector that automatically identifies object positions and classes.Human Pose Estimation — a model for automatic annotation of keypoints for human skeletons.Best suited for: Large volumes of data where annotation needs to be done as quickly as possible, and an annotator only needs to review and fine-tune results.TrackersTrackers are tools for annotating objects across video frames. The annotator marks an object in the first frame, and the tracker automatically follows it across subsequent frames.Models used:TransT — a tracker based on the Transformer architecture for stable multi-object tracking in video.Best suited for: Creating datasets for training tracking models or annotating video quickly.Popular Models for Semi-Automated AnnotationFor Fully Automated Annotation (Detectors)Detectron2 (Facebook AI): A popular framework for object detection and segmentation.EfficientDet: A family of object detection models by Google, ideal for large-scale annotation.CenterNet: Used for keypoint detection and object localization.RetinaNet: A detector focused on challenging scenarios, such as a large number of small objects.For Semi-automated Annotation (Interactors)RITM (Robust Interactive Segmentation Model): Enables mask generation via annotator clicks.F-BRS / DEXTR: Models that refine masks based on keypoints or a bounding box.For Video TrackingDeepSORT: A popular multi-object tracker used extensively in computer vision for following a large number of objects.ByteTrack: A tracker that provides stable results, even with short occlusions.Advantages of AI ToolsTime efficiency: Speeds up annotation thanks to automatic or semi-automatic labeling.Better suited for large datasets: Increases efficiency when dealing with significant data volumes.Integrated workflow: Operates within CVAT itself — no external applications required.Flexibility: Enables a combination of tools for different annotation stages.Disadvantages of AI ToolsNot ideal for complex cases: Heavily overlapping or unclear objects, low image quality, or unusual angles can cause errors.Needs review and adjustment: Fully automatic annotation rarely delivers a perfect result — human verification is still required.Condition-sensitive: Accuracy depends on data quality (lighting, sharpness, object positioning).Not suitable for all data types: Highly specialized datasets may require training custom models.ConclusionAI Tools in CVAT provide a convenient and effective solution for creating computer vision datasets. They help annotators save time and effort but do not fully replace manual work. To achieve maximum precision, combining AI Tools with the annotator’s expertise is essential. When used correctly, AI Tools significantly streamline the annotation process, making it feasible to build high-quality datasets even for large and challenging projects.

AI Tools in CVAT (Overview)

AI Tools in CVAT (Practical Task)

A labeling guideline or specification is a regulatory document that contains a detailed description of the rules, requirements, and criteria governing the data annotation process within a specific project. It ensures a shared understanding of tasks between the client, who defines the goals and expected outcomes, and the annotators, who perform the labeling work.Specification Structure1. Project GoalThe specification helps annotators clearly identify which objects need to be annotated and why. This is a key element of the project context, without it, annotation results may be formal but misaligned with the intended objectives.For example, one project may require annotating people for counting purposes, while another may focus on tracking movement and trajectories.Understanding the project goal reduces the risk of ambiguous decisions, errors, and rework.2. List of Classes and Their DescriptionsEach annotation project specification must include a set of classes, object categories that need to be labeled.Additionally, attributes may be added, extra properties that describe an object’s condition, behavior, or characteristics without introducing new classes.Example: Classes Car and MotorcycleClass: CarDescription:A four-wheeled motor vehicle designed for transporting passengers. Includes sedans, hatchbacks, station wagons, SUVs, crossovers, and coupes.Include:All body typesTaxisCars both moving and parkedExclude:Trucks, busesSpecialized vehicles Attributes:color: red, blue, white, black, gray, othertype: sedan, hatchback, SUV, pickup, wagon, coupelicense_plate_visible: yes, noClass: MotorcycleDescription:A two- or three-wheeled motor vehicle designed to carry one or two people. Includes standard motorcycles and scooters.Include:Motorcycles with or without ridersScooters, mopedsMotorcycles with sidecarsExclude:BicyclesElectric scootersToy or unconventional vehiclesAttributes:color: red, black, blue, white, green, otherrider_present: yes, notype: standard, sport, scooter, off-road, cruiser3. Shape TypesThe specification should define which shapes should be used: bounding box, polygon, polyline, points, etc.Different shapes may be used for different objects within the same project. In some cases, shapes may be combined, for example, a bounding box used together with a skeleton to detect humans and their poses.4. Annotation AccuracyThis is a critical parameter. The higher the required accuracy, the more time-consuming and expensive the annotation becomes. The specification must clearly state:Whether pixel-perfect accuracy is required or if small offsets are acceptableWhether fine details need to be captured (e.g., individual branches of a tree vs. the general outline of the crown)5. Corner Cases and Additional RulesEvery project includes rare, ambiguous, or unusual scenarios (corner cases), that may not be fully predictable. However, common ones should be documented in the specification, and the rest should be identified and clarified during the annotation process.Examples of issues and considerations:Overlapping ObjectsRules for handling partially or fully overlapped objectsAllowable overlap threshold for annotation (e.g., less than 70%)Whether to label only visible parts or reconstruct hidden areasDistant or Small ObjectsMinimum pixel size for objects to be annotatedWhether to annotate all visually recognizable objects, even tiny onesSpecial rules for segmentation tasks, whether to annotate or consider them as backgroundObject InteractionsRules for compound or closely interacting objects E.g., does a bag count as part of a pedestrian or should it be annotated separately?Masks and HolesGuidelines for using masks, including handling holes and gaps E.g., should holes in fences be annotated or left empty?Reflections and Printed ImagesClarify whether to annotate reflections in glass, water, or metalDefine treatment of people and objects depicted on billboards, posters, screens, or packaging By default, these are not annotated unless specified (e.g., for ad or screen content detection)How Annotators Should Work with the SpecificationThe specification is a key project document that defines all annotation rules. The quality of annotation directly depends on how thoroughly the annotator studies and applies the specification. Every section matters, so it’s important to read the entire document to understand all requirements and nuances.Since the document can be lengthy, take time to review it. It’s recommended to first read it carefully as a whole to form a general understanding, then study sections one by one while asking clarifying questions.Below is a step-by-step guide on how to work with the specification:1. Project Goals and Main Requirements (Top Priority)Before you begin, ask yourself the following key questions to understand the context and avoid systematic mistakes:What is the project's goal? Knowing the final goal is the foundation for accurate annotation.Which objects need to be annotated? The specification lists all object classes. Be sure you clearly understand what each class includes and excludes. Descriptions and examples will help with edge cases.Do objects have attributes? In addition to object annotation, you may need to specify extra parameters like color, condition, or movement direction. Study these requirements in advance.Which objects should be excluded? It's equally important to know what not to annotate. For example, there might be ignore zones or minimum object size rules.Which shapes should be used? The specification usually defines the correct shape type for each object: bounding box, polygon, polyline, keypoints, etc. Using the wrong shape may lead to re-annotation.What is the required accuracy level? The specification must state the level of contour precision: whether gaps or overlaps are allowed, and to what extent.2. Additional Details and Special Cases (Second Priority)After understanding the core rules, focus on detailed guidance. This helps when dealing with more complex scenarios:How to handle partially visible or overlapping objects? Should you annotate someone if only part of their body is visible? Should hidden areas be reconstructed or not?How to handle corner cases (rare or debatable situations)? Specifications usually include examples of such cases in a dedicated section.What to do if in doubt?If there are gaps or ambiguities in the document, ask your team lead or project coordinator before you start. This avoids widespread errors.3. Work Through Classes Step-by-StepIf the specification is extensive, after your initial review, you may focus on one class at a time:Select a class and study all rules and nuances carefullyAnnotate only objects from that class on one frame (if there are many) or the entire task (if there are few)Then move on to the next class and repeat the processThis method helps solidify the rules for each class and reduce mistakes.4. Use the Specification During AnnotationThe specification should serve as your working reference at every step:Refer to it whenever you’re unsureCompare your annotations with visual examplesCheck whether you're following all described rules throughout the task5. Quality Control of Your WorkDuring annotation, ask yourself:Am I annotating only the objects listed in the specification?Are my annotations precise and aligned with project requirements?Are there extra or missing objects?Am I handling edge cases correctly based on the guidelines?Self-monitoring helps prevent errors and reduces the need for revisions.6. Feedback and ClarificationsIf you encounter an unclear point or conflict in the specification:Document the question and provide an exampleSend it to the team lead or project coordinator for clarificationNever assume the rules, this can cause systematic errors in the datasetConclusionThe specification is the foundation of a successful annotation project. It defines clear rules and standards that ensure consistency and high-quality data.Understanding the project goal, accurately interpreting classes, knowing shape requirements, and handling corner cases all contribute to effective, error-free annotation.Even with a large specification, a structured approach and class-by-class workflow can reduce the burden and improve comprehension.Always refer back to the document during your work and seek clarification when needed, this guarantees that the annotations meet the expected standards and client goals.Only a responsible and attentive approach to the specification ensures high data quality and the success of the entire project.

Annotation Guidelines: How to Create Labeling Rules

IntroductionData annotation plays a fundamental role in preparing training datasets for machine learning systems. The quality of annotation directly affects the accuracy of models and their applicability in real-world scenarios. Errors at the annotation stage can not only reduce the effectiveness of an algorithm but also distort results entirely.Annotation quality is determined by several factors: completeness, accuracy, consistency of classes and attributes, and adherence to additional rules specified in the project specification. Let’s examine these aspects in more detail.Common Quality IssuesBased on annotation practice, several typical errors can be identified that frequently occur and significantly reduce the value of the data:Missing objects or frames.Crude geometry.Class inconsistencies.Attribute errors.Ignoring additional rules.Each of these issues may seem minor on its own, but together they create systematic distortions: the data becomes incomplete, inconsistent, and fails to reflect reality. Therefore, attention to detail and compliance with all specification rules is a key factor in successful annotation. Let’s look at each point in more detail.Annotation CompletenessThe first and fundamental requirement is completeness. This means that all images included in the task must be processed, and all objects specified in the project specification must be annotated without exception.If even a portion of objects is missed, the dataset becomes incomplete. This leads to distorted model training results: the algorithm cannot "see" all possible object variations, and its predictions become less accurate. For example, if rare or small objects are consistently not annotated, the model will not learn to recognize them.Completeness also requires attention to detail. Small, partially occluded, blurred, or infrequent elements should be annotated along with primary objects. Their absence is often critical, as such examples help the model be more versatile and robust under various conditions.Additionally, annotation completeness involves:Processing all frames/images - no part of the dataset should remain unannotated. Even seemingly "empty" or unimportant frames should be carefully reviewed, otherwise objects present in these frames may be missed.Covering the full range of classes - every object listed in the specification should be annotated. Missing a class creates a data bias and leads to "blind spots" for the model.Consistency among annotators - if multiple annotators work on a project, everyone must interpret what counts as an object in the same way.Thus, completeness is not just about the number of annotated objects but about the integrity and systematic approach to annotation. It ensures the final dataset reflects all aspects of real data and can reliably train a model.Geometric AccuracyAnother critical aspect is geometric accuracy when outlining objects. This determines how closely the shape and boundaries of the annotated object match its actual contour in the image.In some projects, simplified polygons are sufficient. A polygon may roughly follow the object’s outline without tracing every curve or fine detail. This approach allows faster annotation when the main goal is to separate objects from the background rather than to achieve perfect pixel-level accuracy.However, other projects demand much higher precision. Polygons must closely match the object’s true shape, capturing fine curves and intricate details. For instance, when annotating road signs, medical images, or small technical components, coarse polygons are insufficient, even minor inaccuracies in the outline can significantly affect model performance.Therefore, the level of geometric accuracy should always align strictly with the project specification. If approximate boundaries are allowed, there is no need for meticulous outlining. Conversely, if detailed annotation is required, any simplification counts as an error.In other words, geometric accuracy is not always “the more precise, the better.” The key is that the level of detail matches the objectives and tasks of the specific project.Consistency of Classes and AttributesConsistency is equally important. Even perfectly outlined objects lose value if assigned to incorrect categories or if attribute logic is violated.Annotation relies on two main elements:object classes (for example: “car,” “bicycle,” “motorcycle”)attributes (for example: “car color,” “object condition”).If different annotators interpret the same object differently, the dataset becomes inconsistent. For the model, this creates contradictory examples: the same object might be labeled “bicycle” in some cases and “motorcycle” in others. As a result, the algorithm becomes less stable and confused in its predictions.Consistency is particularly critical in cases of:Similar classes - for example, “jaguar” and “leopard,” “sedan” and “hatchback,” ‘bush’ and “tree.” If you don't follow strict specification rules, such objects can easily be confused.Attribute values - for example, when annotating a car’s color, values must be used consistently. If one annotator writes “red”, another writes “blue”, and a third leaves the attribute blank, the resulting data becomes inconsistent.To ensure consistency:Strictly follow the specification - it defines how objects are classified and described.Train and calibrate annotators - so the team shares the same understanding of class and attribute criteria.Review and validate results - validation helps identify discrepancies and correct them before forming the final dataset.Thus, consistency of classes and attributes is fundamental to annotation integrity. Without it, even accurate and complete annotation loses meaning because the data is no longer uniform and cannot train the model effectively.Additional RequirementsBeyond basic rules, many projects include specific constraints tailored to the intended use of the data. These requirements balance completeness and annotation usefulness.Common constraints include:Object size - for example, only annotate objects occupying at least 20×20 pixels. Very small elements may be noise and interfere with model training.Rare categories - in some projects, rare objects are emphasized to provide sufficient examples. In others, they may be excluded if they have low value. For example, in a traffic scene, dozens of cars may appear alongside a single construction vehicle, which may be specially annotated or ignored depending on the task.Annotation conditions - for example, annotate only moving cars and ignore parked ones, or annotate only full-body humans, excluding partially visible pedestrians.Violating these rules makes the dataset inconsistent with project goals. Annotating very small or irrelevant objects may reduce model accuracy on important examples, while ignoring rare classes prevents the algorithm from learning them.Thus, additional requirements act as filters defining the “rules of the game” for annotators. They focus efforts on objects and situations that matter most, excluding irrelevant elements.ConclusionAnnotation quality is the foundation for reliable and effective machine learning models. It encompasses several interrelated aspects:Completeness - all images and objects specified must be annotated. Missing elements, even small or rare, create blind spots and distort data, reducing model accuracy.Geometric accuracy - object contours must meet project requirements, from rough bounding boxes for quick detection to pixel-level segmentation. Incorrect shapes lead to training errors and lower prediction quality.Consistency of classes and attributes - objects must be correctly categorized, and attributes consistently applied. Inconsistencies create contradictory examples that hinder learning and reduce model stability.Compliance with additional requirements - object size, annotation conditions, rare categories, and other rules maintain balance between completeness and utility. Violating these rules renders the dataset unsuitable for project objectives.High-quality annotation ensures the reliability and applicability of a model in real-world conditions. Any errors during labeling propagate into training, reducing algorithm accuracy and stability. Strict adherence to the specification, attention to detail, and consistent annotator practices produce coherent, consistent, and complete datasets.In summary, good annotation is not only careful work by annotators but also the result of well-defined project rules, diligent oversight, and a systematic approach. It minimizes errors, ensures data reliability, and provides a foundation for robust and accurate machine learning models.

What “Good Labels” Look Like

IntroductionHigh-quality data annotation is essential for training reliable machine learning models. Even small mistakes in annotations can negatively affect model performance, especially in large datasets. In practice, annotators may misinterpret instructions, miss objects, or work inconsistently. As projects grow, manually reviewing every annotation becomes difficult and time-consuming.To solve this, CVAT includes built-in Quality Control tools. These tools help identify errors early, maintain consistent annotation quality, and ensure that only reliable data is used in production. In this lecture, we will explore how quality control works in CVAT, focusing on Ground Truth, Honeypots, and key quality assessment settings, as well as best practices for validation and metric interpretation.Manual ValidationWhile automatic validation helps scale quality control and detect errors early, manual validation is still widely used, especially when tasks are complex or labeling rules require human judgment.In practice, manual validation usually follows one of these approaches:Full review (100%): every annotation is checked by a reviewer before it is accepted. This provides the highest level of quality assurance, but is time-consuming and costly.Sampling-based review: only a subset of annotations is reviewed (for example, a percentage of jobs or selected frames). This approach is faster and more scalable, but does not guarantee full dataset quality.Cross-validation: annotators review each other’s work in a rotating or cross-assigned manner to reduce bias and improve consistency. It helps share best practices across the team, but requires clear guidelines to keep feedback objective.These methods are often combined with automatic validation to balance quality, speed, and cost.Automatic validation methodsGround Truth (GT)Ground Truth (GT) is the most universal method of automatic annotation verification in CVAT. It represents a carefully annotated reference dataset that serves as a benchmark for evaluating annotators’ work. GT frames are annotated with maximum accuracy and attention to detail and are then used to compare regular annotations against this “gold standard.”GT can be created at any stage of the annotation workflow and does not interrupt ongoing annotation jobs. Automatic verification using Ground Truth currently supports all 2D annotation types except 2D cuboids.Creating a Ground Truth Job in CVATTo create a Ground Truth job, you start with an already annotated dataset and click the “+” button in the job creation interface. The Job Type is automatically set to Ground Truth. The key configuration step is selecting the Frame Selection Method.Random SelectionWith the Random method, CVAT selects frames randomly from the entire dataset. You can control the size of the validation set in two ways:Quantity – a percentage of the datasetFrame Count – an exact number of framesThese values are synchronized: changing one automatically updates the other. For reliable validation, it is generally recommended to select between 5% and 15% of frames. Smaller datasets usually require a higher percentage to produce meaningful results.A useful option here is the Seed value. Using the same seed guarantees the same frame selection in the future, ensuring reproducibility. Leaving it empty produces a new random selection each time.The main drawback of the Random method is uneven distribution. Some jobs may contain several validation frames, while others may receive none, making it impossible to evaluate quality for those jobs.Random per Job SelectionThe Random per job method addresses this limitation. Instead of selecting frames across the entire dataset, CVAT selects a fixed percentage of frames within each job.For example, if each job contains 20 frames and you set the Quantity per job to 10%, each job will receive exactly two Ground Truth frames. This guarantees balanced coverage and ensures that every job can be validated.Compared to Random selection, Random per job provides more predictable and controlled validation distribution, which is especially useful for structured workflows.HoneypotsConcept and MotivationWith Ground Truth, the number of validation frames grows linearly with the dataset size. In large projects, this becomes costly because GT frames require more careful and time-consuming annotation.Honeypots solve this problem by reusing the same validation frames across multiple jobs. A small set of Ground Truth frames is annotated once and then secretly inserted into regular annotation jobs. Annotators are unaware of which frames are validation frames. After completion, the system compares their annotations with the reference and assigns quality scores automatically.This approach is particularly effective for large-scale projects where scalable quality control is required.Limitations of HoneypotsHoneypots come with several constraints:Supported only for 2D tasks (excluding 2D cuboids)Can be configured only during task creationNot suitable for tracking tasks or workflows requiring strict frame orderIf the validation pool is too small, annotators may start recognizing validation frames, especially when Immediate Feedback is enabledCreating a Task with HoneypotsWhen creating a task, select Honeypots in the Quality section.Key parameters include:Total honeypots – the size of the overall validation set, expressed as a percentage of the datasetOverhead per job – the number of honeypot frames added to each job, relative to job sizeFor example, with 100 frames split into jobs of 20 frames:Setting Total honeypots to 3% creates a validation set of 3 framesSetting Overhead per job to 10% inserts 2 honeypot frames into each jobUnlike standard Ground Truth jobs, the GT job created for Honeypots cannot be deleted. However, individual frames can be removed from the validation set if needed.Frame Distribution and UsageIn Honeypots mode, validation frames are inserted into jobs as additional frames. These frames are ignored during annotation export and are used only for quality evaluation. As a result, job sizes may differ slightly.You can expand the validation pool by increasing the Total honeypots value. A larger validation set allows each job to receive unique honeypot frames, improving the reliability and diversity of quality checks.CVAT also provides statistics on frame usage, allowing you to see how many times each validation frame has been used across jobs.Choosing the Right Validation Set SizeThe reliability of quality metrics depends not on the number of validation images, but on the total number of annotations within those images.If each image contains only one object, quality estimates become coarse and unreliable. In contrast, images with many annotated objects provide more data points and more stable metrics.As a general guideline:Focus on the total number of validation annotations, not just imagesAim for roughly twice the minimum number of annotations needed to measure the desired quality levelFor example, to verify an accuracy of around 80%, approximately 20 validation annotations are recommended.The goal is to balance reliability with efficiency: the validation set should be large enough to produce trustworthy results without unnecessarily increasing annotation cost.Quality Assessment SettingsQuality assessment settings define how CVAT compares annotator work with Ground Truth.General SettingsTasks can inherit quality metrics from the project or use custom settings. The Target metric determines how quality is calculated:Accuracy – overall correctness, including missed and extra objectsPrecision – correctness of annotated objectsRecall – completeness of annotationsThese metrics are based on:True Positives (TP)False Positives (FP)False Negatives (FN)The Target metric threshold sets the minimum acceptable score and is used for Immediate Feedback.Empty Frames HandlingThe Empty frames are annotated option controls whether frames without objects are included in metric calculations. Enabling it produces stricter and more realistic metrics, while disabling it can inflate scores by ignoring empty frames.Job Selection and Validation LimitsFilters allow you to specify which jobs are included in quality calculations. The Max validations per job parameter limits how many correction attempts an annotator can make.Shape, Keypoint, Line, and Group ComparisonWhen evaluating annotations, it is important to check not only whether an object is annotated, but also how accurately it is annotated. CVAT uses several comparison methods, each designed for a specific type of annotation. These methods allow the system to assess accuracy, consistency, and correctness in a clear and reliable way.Shape ComparisonShape Comparison is used for bounding boxes, polygons, and masks.The system compares the annotated shape with the reference shape from the Ground Truth and checks how much they overlap. If the overlap is large, the annotation is considered accurate. If the overlap is small, the annotation is treated as incorrect.Two thresholds control this process:Minimum overlap thresholdThis defines the lowest acceptable overlap. If the annotated shape overlaps the reference shape less than this value, the annotation is considered wrong.Low overlap thresholdThis defines a higher level of overlap. If the overlap is above the minimum threshold but below this level, the annotation is accepted but marked with a warning, indicating that the shape could be more precise.This approach helps distinguish between completely incorrect annotations and those that are mostly correct but need improvement.Keypoint ComparisonKeypoint Comparison is used for point-based annotations, such as landmarks or skeleton joints.The system checks how far each annotated point is from the reference point. An acceptable distance is defined using the OKS sigma parameter.A smaller sigma value makes the check stricter and allows very little deviation.A larger sigma value allows more tolerance and accepts small inaccuracies.For groups of keypoints or skeletons, only visible points are evaluated, and the final score is calculated as an average across all points. The acceptable distance is calculated relative to the object size or, in the case of single points, relative to the image size.This ensures that points are placed close to their correct positions without requiring perfect precision.Point Comparison (Single Points)Point Comparison applies to individual points that do not belong to a group or skeleton.In this case, the matching logic is similar, but the reference size is different. Instead of using a bounding box, the acceptable distance is calculated relative to the image size. This ensures consistent evaluation regardless of where the point appears in the image.This behavior is controlled by the Point size base setting:Group bbox size is used for keypoints and skeletons.Image size should be used for single, standalone points.Line ComparisonLine Comparison is used for polylines, such as roads or boundaries.Instead of requiring an exact match, CVAT creates an allowed area, or envelope, around the reference line. If the annotated line stays within this area, it is considered correct. If it goes outside, the annotation is treated as incorrect.Relative thicknessDefines how far the annotated line can be from the reference line.Controls how strict the line comparison is.Check orientationChecks whether the line direction is correct.Shows a warning if the direction does not match.Min similarity gainDefines how much the match must improve after reversing the line.Used to detect lines drawn in the opposite direction.Group ComparisonGroup Comparison is used to verify logical relationships between objects.When this option is enabled, CVAT evaluates groups as a whole rather than individual objects only. Objects in the annotated group are compared with objects in the reference group.The group match threshold defines how many objects must match for the group to be considered correct. If the number of matching objects is below this threshold, the system reports a group mismatch.This check helps ensure that objects are not only annotated correctly, but also organized into meaningful and correct groups.Segmentation ComparisonFor masks and polygons, CVAT supports:Visibility checks for overlapping objectsMinimum visibility thresholdsOptional comparison of only visible partsThese settings help detect errors caused by incorrect layer ordering or excessive occlusion.Overview and ReportingThe Overview page summarizes annotation quality at task and project levels. It displays:Mean annotation qualityNumbers of correct, missing, and extra annotationsWarnings and errorsJob-level and task-level statisticsQuality reports can be downloaded in JSON or CSV format for further analysis.ConclusionCVAT’s Quality Control tools provide a comprehensive framework for managing annotation quality at scale. Ground Truth and Honeypots address different validation needs, while flexible metrics and comparison settings allow teams to tailor quality checks to their workflows.By carefully selecting validation strategies, configuring metrics appropriately, and interpreting results correctly, teams can ensure consistent, high-quality annotations while keeping annotation costs under control.

Quality Control Methods in CVAT