What initially attracted me to CVAT.ai was the ability to use the free cloud version to learn the platform and validate my annotation workflow before making any investment. That allowed me to develop my annotation schema, understand the tool's capabilities, and confirm it met the needs of my research. As the scope of my doctoral project grew, I discovered that CVAT also offers Professional Labeling Services, which made it possible to scale the annotation effort while maintaining the quality standards I had established. The biggest strength of CVAT.ai was the combination of a flexible annotation platform and a collaborative Professional Services team. My research required far more than basic bounding boxes. I developed a custom annotation schema with five object classes and additional attributes for blur, occlusion, and truncation, along with detailed annotation guidelines and objective quality acceptance criteria. The Professional Services team worked through multiple review cycles, incorporated feedback as the annotation guidelines evolved, and consistently applied the specification across the dataset. The final dataset integrated directly into my YOLO training pipeline and became the foundation of my doctoral research. The ability to move seamlessly from evaluating the platform on my own to partnering with the Professional Services team made CVAT a valuable part of the project from start to finish.

I like pre-annotating data with CVAT.ai, it works really well. The platform is intuitive to use, and I quickly found what we intended to do. Compared to our self-built tool, CVAT.ai is simply incomparable. It greatly reduces the labeling effort, I estimate to a quarter to a fifth of what it would otherwise be. Overall, I find it really good, even if it might take a bit of time to get used to.

The biggest win for us is the model-assisted annotation loop. We run YOLO and SAM2 pre-annotation directly through CVAT, so the work is mostly verifying and correcting rather than labeling from scratch. For a small operation producing the computer-vision training data behind our product, that's the difference between a dataset taking weeks versus days. We also rely on the flexibility of handling multiple annotation types in one tool — we annotate ball and equipment bounding boxes, hockey stick parts, and pose keypoints, and the project/task structure keeps dozens of video tasks organized without friction. Being cloud-hosted means there's no annotation infrastructure for us to stand up or maintain, and the export formats drop straight into our YOLO/training pipeline.