Accuracy - prnvkhndlwl/Kinetrace GitHub Wiki
Learned point trackers see the video at reduced resolution (AllTracker at up to 1024 px, CoTracker3 at ~512×384 internally), so their raw localization error grows with resolution: CoTracker3 measured ≈3.7 px mean on 4K synthetic ground truth. This tool therefore runs a second stage at native resolution — a gated Lucas-Kanade refinement seeded from the model's own prediction, chained from your exact seed click and re-anchored to the model every frame (it can never drift more than a few pixels from the model's track). Measured result on 4K ground truth: ≈0.7 px mean / 1.8 px max. At ≤720p the model is already at native scale and refinement automatically stays out of the way. ROI zoom attacks the same root cause from the other side: inside a crop, small objects reach the model at several times the detail, which is what makes tracking robust to background changes around them.
Tips for best precision and robustness:
- Click points on visually distinct features when you can — or better, drag a circle snugly around the object so it tracks as a region.
- The seed frame keeps your exact clicked coordinates in the export — corrections are treated as ground truth.
- Track a few extra "sacrificial" points on the same rigid object: the point model tracks all points jointly, so they support each other. Ignore the extras at export (or just use a region, which does this internally).
- Trust the timeline reds: they mark exactly where the model struggled.