Yifu Zhang

It seems that you do not compile DCNv2 successfully. You should compile DCNv2 first.

Only MOT15 and MOT20 need to be finetuned. You can find the traininig file here: https://github.com/ifzhang/FairMOT/blob/master/experiments/ft_mot15_dla34.sh

You can use ResNet-34, which is much faster. The sh file of training ResNet-34 can be found in the experiments folder.

Yes, max is more reasonable. When using sum, we need to add a clip option to keep the value between 0 and 1. In fact, the code is using max...

We give each bbox a unique id and the total ids are larger than 1,000,000.

You can check whether you load the coco pretrained model correctly. When loading coco pretrained model and trained on half MOT17, you can get 67+ MOTA. If you train from...

mot17.json contains all the training images in mot17 and mot17_half.json contains half of training images of mot17 (for each sequence, first half training, second half evaluation).

It seems that it's a reid failure. conf_thres helps get better detection results. If you want to modify the reid results, you can go to multitracker.py and there are some...

You can change the format of your custom dataset and make it the same as MOTChallenge dataset because the evaluation code only support the MOT format dataset.

Yes, we can only evaluate on MOT16-train. If we want to evaluate MOT16-test, we need to submit the results to the MOT Challenge website.