The open-source label factory behind half the vision datasets you've seen

What it does

CVAT Community is a browser-based platform for labeling images, videos, and 3D point clouds. Teams create projects, assign annotation jobs, review each other’s work, and export to 20+ formats including COCO, YOLO, and KITTI. It runs entirely on your own infrastructure via Docker Compose, with a Python SDK and REST API for pipeline automation.

The interesting bit

The project treats your own ML models as first-class citizens: you can wire in detectors, segmenters, or trackers (SAM, YOLO, RetinaNet, etc.) through a serverless Nuclio layer to pre-label data before humans touch it. That’s a pragmatic inversion of the usual SaaS pattern where AI assistance is a premium upsell.

Key highlights

• MIT-licensed core with active maintenance since 2018
• Multi-organization support with role-based access and review workflows
• 20+ import/export formats plus direct S3/Azure/GCS storage hooks
• Pre-built serverless integrations for PyTorch, ONNX, OpenVINO, and TensorFlow models
• Grafana dashboards for tracking annotator activity and project progress
• Chromium-based browsers only; Firefox has caveats, Safari is unsupported

Caveats

• Advanced analytics, SSO, SAM 2/3, and AI agents are paywalled to CVAT Online or Enterprise tiers
• Some serverless model assets carry separate non-commercial licenses despite the MIT core
• FFmpeg dependencies pull in LGPL/GPL obligations you’ll need to track

Verdict

Worth evaluating if you need to keep data in-house and already have models to deploy against your labeling queue. Skip if you want turnkey SaaS convenience or need Safari support for your annotation team.