search

ifzhang / ByteTrack

[ECCV 2022] ByteTrack: Multi-Object Tracking by Associating Every Detection Box
MIT License
4.8k stars 906 forks source link

BYTETrack evaluation changes thresholds per sequence? #144

Open agirbau opened 2 years ago

agirbau commented 2 years ago

Hello!

First, thanks for your great work, it's amazing! I was wondering around the repo, and I found this piece of code on mot_evaluator.py, which seems to be changing thresholds and buffers depending on the sequence it's evaluating. Is this correct? I was wondering if these numbers follow some pattern over the different sequences (e.g. MOT-05 and 06 being sequences with a of camera motion). Thanks!

https://github.com/ifzhang/ByteTrack/blob/4829efd6ffc2882fd4e0fe7ea5457e90e36e898e/yolox/evaluators/mot_evaluator.py#L138-L160

Ali96528 commented 1 year ago

Hello!

First, thanks for your great work, it's amazing! I was wondering around the repo, and I found this piece of code on mot_evaluator.py, which seems to be changing thresholds and buffers depending on the sequence it's evaluating. Is this correct? I was wondering if these numbers follow some pattern over the different sequences (e.g. MOT-05 and 06 being sequences with a of camera motion). Thanks!

https://github.com/ifzhang/ByteTrack/blob/4829efd6ffc2882fd4e0fe7ea5457e90e36e898e/yolox/evaluators/mot_evaluator.py#L138-L160

hello,agirbau Have you figured out why using different track_buffer and thresh in different sequences ? I'm curious about it ,too. Please let me know ,if you know the answer. Thank you!

agirbau commented 1 year ago

My guess is that they tested several thresholds/hyperparameters to have the best results on the test sequences.

This CVPR paper seems to suggest the same in Table 3 "... since ByteTrack uses different thresholds for different sequences of the test set and interpolation we recomputed their results...".

Ali96528 commented 1 year ago

My guess is that they tested several thresholds/hyperparameters to have the best results on the test sequences.

This CVPR paper seems to suggest the same in Table 3 "... since ByteTrack uses different thresholds for different sequences of the test set and interpolation we recomputed their results...".

Thanks for the reminder.