search

Zhongdao / Towards-Realtime-MOT

Joint Detection and Embedding for fast multi-object tracking
MIT License
2.38k stars 539 forks source link

Different evaluation results on MOT16 training set #170

Open iriswang02 opened 4 years ago

iriswang02 commented 4 years ago

Hi, @Zhongdao

Thank you for your code.

I have successfully installed your repo together with essential requirements. It runs very well but I just want to check one specific detail with you. Hope you can give some advice.

The performance results you shared in README.md file is tested on the MOT-16 training set by using Nvidia Titan Xp GPU.

I do exactly what you did but get a slightly different result. For example, using JDE-1088x608, I get 73.1 on MOTA instead of 74.8. I'm not sure whether this small difference comes from the different GPU, may I know your opinion on this? And is it normal?

Thank you for your great help!

IwenLeeO commented 4 years ago

Hi @iriswang02 I have some questions when training the model, could you please tell me how to train the model to get 73.1using JDE -1088x608. i always get 54% .Thanks

nemonameless commented 3 years ago

@Zhongdao same question 73.1 on MOTA instead of 74.8

Moens007 commented 3 years ago

Hi @iriswang02 I have some questions when training the model, could you please tell me how to train the model to get 73.1using JDE -1088x608. i always get 54% .Thanks

yes,I am always this score and I do not know why, maybe the numbers of GPU or batchsize cause? i want to know your parameters about it.