CVAT setup and bounding box tracking tutorial

This CVAT tutorial covers browser-based setup and bounding box annotation for video datasets. The speaker creates an account with a working email, confirms it through an email link, then creates a project. They describe upload limits on the testing account and say larger client projects require paid membership; the walkthrough gives no specific quota or price.

Project setup uses the label constructor to define rectangle classes and colors. The speaker then creates a task, uploads a dataset and selects frames 0 through 99 for a 100-frame example. They recommend 70% image quality to reduce lag, while noting that a client may require 100% quality.

The annotation section explains two-point rectangles and the difference between individual shapes and tracks. In the demonstration, N repeats the previous drawing action, F advances one frame and D returns to the previous frame. Tracking keeps a box across frames, with manual adjustments at keyframes and interpolation between them. The speaker suggests checking boxes every five frames for slow movement or ten for very slow movement, then inspects intermediate frames and corrects misplaced edges.

The tutorial also explains why a client may require one object ID throughout a track. Separate manual shapes can have different IDs. It ends with deleting boxes, changing classes, undoing edits and copying objects. The demonstrated workflow is manual annotation and tracking; the speaker says they rarely use CVAT's AI tool.