The precision about pymdnet

[zhangyujia0223]

@Handworker @HyeonseobNam I am just beginner of tracking evaluation. Could you show me how to do evaluation ? Thanks very much!

[MalFan]

I also found that the result is far lower than the matlab version.

[Zealoe]

have you found the reason? thanks

[xpain1]

@Handworker @MalFan I also found that the result is far lower than the matlab version.What is your precision on OTB-100 dataset when you evaluate?

[sydney0zq]

@Handworker Hey, do you finally find out the reason of precision? I could see the owner of this repostority is also the author of the MDNet paper. I am confused whether it is so good as the matlab version.

[jimmy-dq]

@sydney0zq The precision of this version is a bit lower than the matlab version (OTB-100: DP-20: 88.89% VS 90.9%). I do not find the exact reason for this problem, but may be caused by the difference between two deep-learning frameworks.

[sydney0zq]

@Handworker Thank you for your timely reply. Recently I deploy py-MDNet on other real video datasets. But the result seems not so good, at least not as good as the demo videos on youtube. I think maybe MDNet is a little overfitting on these popular benchmarks. And more, if we want to reproduce the MATLAB version result, we could convert released matlab convnet model into pytorch state dict, which is very easy and feasible. Anyway, thank you for your reply.

[JellyWong]

@sydney0zq Actually, I have checked the parameters of the released pytorch model and I found that the pytorch model indeed shares the same weights with the matlab version. I assume that the authors just lossless transfer the matlab pretrained model to the current pytorch one.

[laisimiao]

@sydney0zq How to convert released matlab convnet model into pytorch state dict? Could you share your transfer codes with me?