The previous evaluation code has a bug that it uses unscaled bbox coordinates to extract features for identity matching. It results in lower accuracy when tracking head does not use additional cues for matching. When additional cues (IoU, category, detection score) are used, the performance does not change much.
Changes
commit 4e92e86200905f24ea37b1572a50f7eb49c5a7d3
Tests
Tested with the current model
Misc
Due to this bug, some results in the arXiv paper is also changed. We will update the results in the arXiv paper shortly. Sorry for any inconvenience caused by this issue.
Background
The previous evaluation code has a bug that it uses unscaled bbox coordinates to extract features for identity matching. It results in lower accuracy when tracking head does not use additional cues for matching. When additional cues (IoU, category, detection score) are used, the performance does not change much.
Changes
commit 4e92e86200905f24ea37b1572a50f7eb49c5a7d3
Tests
Tested with the current model
Misc
Due to this bug, some results in the arXiv paper is also changed. We will update the results in the arXiv paper shortly. Sorry for any inconvenience caused by this issue.