CVAT — Computer Vision Annotation Tool for AI Training Data

Introduction

CVAT (Computer Vision Annotation Tool) is an open-source platform for labeling images and videos to create training datasets for machine learning models. Originally developed by Intel, it provides a web-based interface where annotators can draw bounding boxes, polygons, polylines, keypoints, and cuboids. It supports team workflows with task assignment, quality control, and export to common dataset formats.

What CVAT Does

• Annotates images and video frames with multiple label types including bounding boxes, polygons, and keypoints
• Supports semi-automatic annotation using AI models via the Nuclio serverless framework
• Manages multi-user projects with task assignment, review workflows, and quality analytics
• Exports datasets in COCO, Pascal VOC, YOLO, Datumaro, and other popular formats
• Provides a REST API and Python SDK for programmatic task creation and data retrieval

Architecture Overview

CVAT runs as a set of Docker containers: a Django backend, a PostgreSQL database, a Redis queue, and an Nginx reverse proxy. The frontend is a React application that communicates with the backend via REST APIs. For AI-assisted annotation, CVAT integrates with Nuclio to deploy inference models as serverless functions that run alongside the main application.

Self-Hosting & Configuration

• Requires Docker and Docker Compose on a Linux host
• Minimum recommended: 4 CPU cores, 8 GB RAM for small teams
• GPU support is optional, used only for AI-assisted annotation models
• Configure SMTP, storage backends, and auth providers via environment variables
• Persistent data is stored in Docker volumes for the database and uploaded media

Key Features

• AI-assisted annotation with automatic bounding box and polygon suggestions
• Built-in analytics dashboard for tracking annotation progress and quality
• Supports both image and video annotation with frame-level interpolation
• Cloud storage integration with AWS S3, Google Cloud Storage, and Azure Blob
• Active Directory and LDAP integration for enterprise single sign-on

Comparison with Similar Tools

• Label Studio — more general-purpose (text, audio, images); CVAT specializes in computer vision
• doccano — focused on text annotation; CVAT handles images and video
• Supervisely — commercial platform with a free tier; CVAT is fully open source
• VOTT — archived by Microsoft; CVAT is actively maintained with frequent releases

FAQ

Q: Can CVAT handle video annotation? A: Yes. CVAT supports frame-by-frame video annotation with object tracking and interpolation between keyframes.

Q: Does it support automatic annotation? A: Yes. You can deploy pre-trained models (such as YOLO or Faster R-CNN) via the Nuclio integration to generate automatic annotations that annotators can then refine.

Q: What export formats are supported? A: CVAT exports to COCO JSON, Pascal VOC XML, YOLO TXT, Datumaro, LabelMe, and several other formats.

Q: Is there a cloud-hosted version? A: Yes. The team offers a managed cloud service at app.cvat.ai with free and paid plans.

Sources

• https://github.com/cvat-ai/cvat
• https://docs.cvat.ai