[Feature] Support multi-class tracking in ByteTrack

[noahcao]

add multi-class multi-object tracking support in ByteTrack based on the labels segments given in detections.

It would be trivial to extend this feature to other trackers.

[codecov]

Codecov Report

Merging #548 (f2be737) into master (88f92dd) will increase coverage by 0.07%.
The diff coverage is 100.00%.

@@            Coverage Diff             @@
##           master     #548      +/-   ##
==========================================
+ Coverage   73.08%   73.16%   +0.07%     
==========================================
  Files         129      129              
  Lines        8019     8065      +46     
  Branches     1562     1563       +1     
==========================================
+ Hits         5861     5901      +40     
- Misses       1664     1670       +6     
  Partials      494      494              

Flag	Coverage Δ
unittests	73.10% <100.00%> (+0.07%)	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.

Impacted Files	Coverage Δ
mmtrack/models/trackers/byte_tracker.py	90.22% <100.00%> (+0.70%)	⬆️
mmtrack/datasets/sot_train_dataset.py	82.82% <0.00%> (-2.03%)	⬇️
mmtrack/datasets/pipelines/transforms.py	86.92% <0.00%> (-1.84%)	⬇️
mmtrack/models/track_heads/siamese_rpn_head.py	93.15% <0.00%> (+0.10%)	⬆️
mmtrack/models/reid/linear_reid_head.py	89.85% <0.00%> (+0.14%)	⬆️
...track/models/trackers/quasi_dense_embed_tracker.py	75.71% <0.00%> (+0.17%)	⬆️
mmtrack/datasets/youtube_vis_dataset.py	77.23% <0.00%> (+0.18%)	⬆️
mmtrack/models/track_heads/roi_track_head.py	87.93% <0.00%> (+0.21%)	⬆️
mmtrack/models/track_heads/roi_embed_head.py	85.34% <0.00%> (+0.25%)	⬆️
... and 10 more

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 88f92dd...f2be737. Read the comment docs.

[GT9505]

1. Does the 'multi-class supporting` be the same as ByteTrack in the manuscript?
2. It is better to verify the inference performance of ByteTrack in MOT17-half-train using the pertrained weights. It should achieve 67.7 HOTA

[noahcao]

1. this 'multi-class supporting' does no require any modification in the manuscript as it is directly based on the labels which is already input to the tracker in the previous version.
2. I have verified the performance alignment before and after this commit. The performance is well aligned though it is not 67.7 HOTA but slightly lower as shown below.

[GT9505]

1. this 'multi-class supporting' does no require any modification in the manuscript as it is directly based on the labels which is already input to the tracker in the previous version.
2. I have verified the performance alignment before and after this commit. The performance is well aligned though it is not 67.7 HOTA but slightly lower as shown below.

Do you mean that we can achieve 67.7 HOTA before this pr, while achieve 67.1 HOTA after this pr?
I do not get it. The behavior should have be unchanged when there is only one category to track, since all element in the cate_cost equal to 0.

[noahcao]

1. this 'multi-class supporting' does no require any modification in the manuscript as it is directly based on the labels which is already input to the tracker in the previous version.
2. I have verified the performance alignment before and after this commit. The performance is well aligned though it is not 67.7 HOTA but slightly lower as shown below.

Do you mean that we can achieve 67.7 HOTA before this pr, while achieve 67.1 HOTA after this pr? I do not get it. The behavior should have be unchanged when there is only one category to track, since all element in the cate_cost equal to 0.

I meant: when I ran the original code (before this PR), it achieved HOTA=67.1. After PR, it is still 67.1. I didn't ever reproduce the reported 67.7HOTA on my machine at the very beginning.

[GT9505]

I'll check it later since the bytetracker is not changed.
I guess that it may due to the changes of yolox detector in mmdet.

[GT9505]

I'll check it later since the bytetracker is not changed. I guess that it may due to the changes of yolox detector in mmdet.

using mmdet=2.19.1 and mmtrack=0.13.0 we can achieve 67.7 HOTA.
using mmdet=2.24.1 and mmtrack=0.13.0 we can achieve 67.1 HOTA.