results on MOT17 validation set in Table 7

[liuqk3]

Hi Yifu,

Thanks for your great work. Here I have an issue about the results in Table 7 (in the newer version paper).

The first row shows the results when only trained on MOT17 is 67.5% MOTA and 69.9% IDF1. I tried to reproduce these results. I trained a model using python train.py mot --exp_id mot17_half_dla34 --load_model '../models/ctdet_coco_dla_2x.pth' --data_cfg '../src/lib/cfg/mot17_half.json' , and test the model_30.pth using python track_half.py mot --load_model ../exp/mot/mot17_half_dla34/model_30.pth --conf_thres 0.4 --val_mot17 True, the results are 60.7% MOTA, 67.3% IDF1, which are much lower than those presented in Table 7.

Do you have any suggestions about the results? Thanks @ifzhang

[ifzhang]

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 scratch, you can get only 60+ MOTA.

[yangyangkiki]

Hi Yifu, I have the same issue with @liuqk3. I try to reproduce the results of the first row in Table7.

I trained the model by using : python train.py mot --exp_id mot17_dla34 --load_model '../models/ctdet_coco_dla_2x.pth' --data_cfg '../src/lib/cfg/mot17.json' --gpus 0,1 , and then test the model_30.pth by using : python track.py mot --load_model ../exp/mot/mot17_dla34/model_30.pth --conf_thres 0.4 --test_mot17 True .

Here is my problem: I found @liuqk3 adopted the mot17_half.json file to train the model, but I choose mot17.json. I read your paper again, and I don't know the difference between these two files. Can you give me some suggestions?

Thank you so much. @ifzhang

[ifzhang]

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).

[Wenbin94]

Have you solved this problem? My test results are similar to yours. MOTA 60.7,IDF1 69.1