It doesn't work well

[FearlessArmy]

trafficstars

According to your file configuration, you can't run the effect as you said, the result is as follows, two GPUs have been used, and the following result is the effect of one GPU python projects/SGT/train_net.py --config-file projects/SGT/configs/MOT17/sgt_dla34.yaml --data-dir /root/datasets --num-gpus 1 OUTPUT_DIR /root/sgt_output/mot17_val/dla34_mot17-CH

[FearlessArmy]

@HYUNJS

[HYUNJS]

That configuration trains SGT using half of the MOT17 training set and evaluates the other half. Actually, all experiments for the paper are conducted using the docker file with CUDA10.1.

Since the following MOT17 split for the ablation experiment is quite small, I also observed ~1 MOTA fluctuation. If you use MOT17 full-set training or joint training with CrowdHuman, such performance gap would be reduced.

[FearlessArmy]

@HYUNJS I also trained with the crowdhuman dataset and the results were just as poor，The best MOTA is 72.8%

[HYUNJS]

Can you try these pretrained weights?

mot17half mot17half+CH