CVAT (Computer Vision Annotation Tool)

What CVAT is

CVAT, the Computer Vision Annotation Tool, is an open-source platform for labelling visual data so machine learning models can learn from it. It started inside Intel and is now developed by CVAT.ai, with the code on GitHub under the MIT licence. It handles images, video, audio and 3D point clouds, and it is built for teams: tasks are split into jobs, jobs are assigned to annotators, and reviewers accept or reject the work.

What you can label with it

• Bounding boxes around objects, the most common computer vision task.
• Polygons and masks for pixel-accurate outlines (segmentation).
• Polylines and points for lanes, edges and landmarks.
• Skeletons for human pose: joints connected in a fixed structure.
• Tracks that follow one object across many video frames, with interpolation between the frames you key.
• Tags and attributes on whole images or on each shape, such as "occluded" or "truncated".

Why it matters for AI data work

Many annotation projects, on vendor platforms and on freelance marketplaces, use CVAT or a tool that works the same way. Learning it on your own data teaches you the habits paid work checks: tight boxes, consistent class names, attributes filled in every time, and keyboard shortcuts so you can work at a steady pace without cutting corners. A small, well-labelled dataset with a short written guide is also a portfolio piece you can show a client.

Getting started

1. Create a free account on the hosted service, or run it yourself with Docker by following the self-hosting guide in the documentation.
2. Create a project, define your labels and their attributes before you start.
3. Upload 50 to 100 images, label them, then label 20 of them again a week later and compare.
4. Export in COCO or YOLO format and open the file to see what a model actually receives.

Free and paid

The open-source version is free to run on your own machine or server. The hosted cvat.ai service has a free tier and paid plans for teams; check the official pricing page, because plans change.

Common questions

Do I need to code? No. Labelling happens in the browser. Python helps only if you want to use its SDK or command-line tools.

CVAT or Label Studio? CVAT is stronger for video, tracking and 3D. Label Studio covers text, audio and documents more broadly. Learning one makes the other easy.

Best for: anyone preparing for image or video annotation work, and students building a vision project on their own data.