search

ifzhang / FairMOT

[IJCV-2021] FairMOT: On the Fairness of Detection and Re-Identification in Multi-Object Tracking
MIT License
3.98k stars 936 forks source link

Instructions for running other models #9

Open niskov opened 4 years ago

niskov commented 4 years ago

First - great project! I followed the instructions and it worked (almost) immediately (had some issues because i'm running it on Windows).

Can you provide instructions for running the demo on the other included models? (ctdet_coco_dla_2x / hrnetv2_w18_imagenet_pretrained / hrnetv2_w32_imagenet_pretrained)

I tried using them but received torch warnings/errors. for some of the files I got 0 detections, for the others it simply didn't work

Thanks in advance

ifzhang commented 4 years ago

These models are just pretrained models. You need to train on the MOT dataset and then run the demo. I will provide my trained HRNet model some days later for you to run the demo which do not use DCNv2. Now only the baseline model all_dla34.pth can run demo.

muzishen commented 4 years ago

@ifzhang Thank you very much for your excellent work! Please, can you tell me baseline model (DLA-34 and HRNetV2_W18) training on which datasets?

ifzhang commented 4 years ago

The training data including Caltech Pedestrian, CityPersons, CUHK-SYSU, PRW, ETHZ, MOT17, following the work of "Towards Real-Time Multi-Object Tracking"

muzishen commented 4 years ago

Wow, do you mean baseline model is trained on all the above training data without MOT15 and MOT20?

ifzhang commented 4 years ago

Yes.

mainguyenanhvu commented 3 years ago

@ifzhang Do you mean that if I want to use pre-trained model *dla, I need to train them on MOT dataset. I found that there is 0 detection when I run demo on these models.

Thank you.