Cvat

CVAT's moat isn't the annotation canvas—Label Studio and Roboflow both draw boxes competently. It's the combination of three things that's hard to replicate: 3D point cloud detection and tracking (.pcd, .bin), skeleton graph tracking across video frames for pose estimation, and a QA layer built around validation/acceptance stages, ground truth jobs, honeypots, and consensus scoring. Teams labeling for autonomous driving, robotics, or gait/motion analysis get those out of the box rather than by writing plugins. Automation is the second differentiator. SAM 2 and SAM 3 handle interactive image segmentation, SAM 2 also covers video segmentation and tracking, and Ultralytics, Hugging Face, and Roboflow models plug in for batch auto-labeling. You can bring your own model through the API. That means a labeling loop where AI proposes masks and humans fix edges, which typically cuts mask-annotation time substantially versus clicking every polygon vertex. The licensing model is the third differentiator and also the biggest decision. CVAT Community is open source and free to self-host—genuinely attractive for research labs and companies with existing Kubernetes infrastructure. It's the same annotation engine as the paid tiers, so you don't trade features for cost; you trade operational burden. Someone has to keep it upgraded, backed up, and secured. On CVAT Online, the paid tiers share a quirk worth understanding before you commit: Solo is explicitly a personal workspace, unlimited but solo. It's not a 'cheap team plan'. The minute a second person needs to review an annotation, you move into Team pricing at $33/user/month monthly or $23/user/month billed annually, and Team enforces a minimum seat count. Multi-stage QA features—manual review, ground truth jobs, honeypots, reports—live on Team and above, which is deliberate: the features that make annotation trustworthy are the ones you pay for. Where CVAT is genuinely weaker: audio is announced but not shipped, so multimodal text+audio+image pipelines still need a second tool. The Free tier is generous enough to evaluate (1 project, 3 tasks, 100 AI agent calls/month) but not to run a real project. And while exports cover 20+ formats, the labeling service is a separate commercial motion—you're either a platform customer or a service customer. The practical stack: evaluate on Free, run the open-source Community edition on your own hardware if you have ops capacity, and step to Team paid annually only when collaboration and QA become real. Enterprises in healthcare, defense, or government with air-gapped or on-prem requirements are the natural buyer for the $12,000/year Enterprise tier—that's where SSO, RBAC, audit logs, and encryption at rest live.