search

researchmm / SiamDW

[CVPR'19 Oral] Deeper and Wider Siamese Networks for Real-Time Visual Tracking
http://openaccess.thecvf.com/content_CVPR_2019/html/Zhang_Deeper_and_Wider_Siamese_Networks_for_Real-Time_Visual_Tracking_CVPR_2019_paper.html
MIT License
751 stars 180 forks source link

Can't get 65+ with GOT10k #86

Open ISosnovik opened 4 years ago

ISosnovik commented 4 years ago

Hello

Thanks for your code. I followed your code and trained ResNet22 on GOT10k. No modification were done. At the end I got 64.7 on OTB15 How did you get 65.4?

JudasDie commented 4 years ago

Hello

Thanks for your code. I followed your code and trained ResNet22 on GOT10k. No modification were done. At the end I got 64.7 on OTB15 How did you get 65.4?

Hi, thanks for your interest. Have you tuned the parameters (see tuning toolkit) ? In my experience, training on GOT10K is easy to get good results.

ISosnovik commented 4 years ago

I used the ones you've provided. Do I need to tune them each time I train my model? So for each set of weights I have a different set of hyperparameters?

JudasDie commented 4 years ago

I used the ones you've provided. Do I need to tune them each time I train my model? So for each set of weights I have a different set of hyperparameters?

I would tune it only if I get a nice result after epoch testing. If you have enough resources, of course you could tune it every time. For example, if your current is 64.7, it should be 65.5-67 in my experience.

davinca commented 4 years ago

@ISosnovik @zhangliliang Have u trained the siamrpn model successfully with GOT10K? I got an situation like this: image my torch version is: torch.1.2.0

ISosnovik commented 4 years ago

@davinca you can check https://github.com/researchmm/SiamDW/issues/38

davinca commented 4 years ago

@ISosnovik thanks!

davinca commented 4 years ago

@ISosnovik @JudasDie I trained ResNet22 on GOT10k and get 47.8 on OTB15,No modification in SiamRPN.ymal. the result is abnormal and confusing.=-=

JudasDie commented 4 years ago

@ISosnovik @JudasDie I trained ResNet22 on GOT10k and get 47.8 on OTB15,No modification in SiamRPN.ymal. the result is abnormal and confusing.=-=

The models provided on readme (got10k) are for "FC" versions (Res22FC and Res22W FC). If you want to use "RPN", please use default configs. If you want to train on GOT10K (as you mentioned), please use FC models (20w pairs each epoch are enough). When you tune the hyper-parameters, I would suggest you start with VOT, since OTB15 containing more frames takes much longer time.

davinca commented 4 years ago

@JudasDie I use FC model to initial the SiamRPN and increase pairs for epoch(40w), but gives me 0.53 on OTB100.

ISosnovik commented 4 years ago

And what is 20w and 40w? @davinca

davinca commented 4 years ago

@ISosnovik it indicates the number of training samples per epoch.

ISosnovik commented 4 years ago

What is w?

ISosnovik commented 4 years ago

Oh. I suppose it is 10 000

davinca commented 4 years ago

yep =-=