About the trainning with ByteTrack_ReID model

[zengjie617789]

Thank you the awesome contributions. Here are some questions about the respository below:

1. Is it resonable to add ReID branch in YOLOx model? the YOLOx has 3 hierachies which of downsample ratio is 8, 16, 32. From my understanding the larger downsample ratio is, the much uncertainty ID features we got. ~~2. About the nID when trainning crowdhuman datasets. The original Fairmot make the output class equel to the total num ID in datasets which is a large num. The trainning of ReID is hard to control. Furthermore, the performance of using ID features to match is worse than using detect results merely. Can you share some evaluated results.~~
2. I am trainning the model on MOT20 datasets generated by convert_mot20_to_coco.py. After that, when I started the tranning I meet a error about loss backward. I finally resolved it by incresed the total_id num like that below: total_ids = max(max_id_each_img) + 1 + 1 # TODO Need Check: ids start with 0 Though it can run successfully, it is curious why should increase the id num by 2 instead 1. Besides, The original Fairmot did the same operations.
3. Not using ID features in demo_track.py. ~~After trainning the model on crowdhuman datasets, I test a video with demo_track.py . After that, I find the tracker is ByteTracker which is not realted to the ID features. So what the right way to test custom video on the trained model with ID features~~ Just replace the Bytetrack with Bytetrack_fairmot and modified some code. but the results i got is not as goog as i expected. I wiil be appreciate if anyone who can help me, Thank you in advance.