search

Zhongdao / Towards-Realtime-MOT

Joint Detection and Embedding for fast multi-object tracking
MIT License
2.37k stars 539 forks source link

Concerns about the datasets used for training the models #168

Closed mauro11-sant closed 3 years ago

mauro11-sant commented 4 years ago

First of all, congratulations on the work. I have been making modifications to the code and I have some concerns about the datasets used in the training.

My doubt is if the provided models (JDE-1088x608, JDE-864x480, JDE-576x320) are trained with all the datasets of the DATASET_ZOO.md file except the MOT16 for validation.

If so, that would be a problem since the MOT17 and MOT16 videos are the same ones, so the training and validation would overlap.

I assume that the MOT17 is not part of the training, but I would like someone to confirm this if you would be so kind.

Thank you for your response. Keep up the good work.

Zhongdao commented 3 years ago

Yes, MOT17 is used in training. For validation, we use MOT15 (which has no overlap videos with MOT17) in the paper. In readme.md I provide the tracking results on MOT16, that's not for validation, just for reference.