The MOT17 benchmark results in Figure 4 of the paper.

[lzzppp]

Thank you for your contribution to the open source community. I encountered some problems when conducting my thesis experiments. I arranged the Coltrack experiment according to the training steps in the ReadMe. I performed pre-training based on the configuration file "single_class.py" in the TBD folder in the config directory. I trained for 30 epochs, and then used this weight to train the MOT task. The training configuration file used the E2E folder. "e2e_submit_mot17.py" configuration file. After 60 epochs of training, I used the best round results and the last round results to compare with ByteTrack. The configuration file used was "e2e_submit_mot17.py" in the E2E folder, but two lines were added: just_inference_flag = True and num_classes = 91. Under the frame interval of "1 2 3 6 10 15 30", the results of coltrack did not reach the results reported in the subgraph of Figure 4(a) of the paper. I followed the suggestions given by ByteTrack to modify the three parameters of the configuration file: score_thresh, filter_score_thresh and miss_tolerance, but it did not help.

[lzzppp]

The following is my experimental environment: NVIDIA RTX A6000*8 Driver Version: 535.129.03 CUDA Version: 12.2

[lzzppp]

Training log [Uploading info.txt.zip…]()

[lzzppp]

Inference settings: I saved the val data set of mot17 to a directory in demo format at different intervals. The frame id was remapped, and they are all in the format of 1 2 3 4..., so the inference_sampler_interval is set to 1 during inference.

[lzzppp]

Below are the experimental results of adjusting several key inference parameters.

[lzzppp]

This is the losses:

[lzzppp]

The HOTA metric in the training:

[lzzppp]

I have fixed my error