Accuracy issues

[HJ-Xu]

Hi, thanks for your great work. But when I train the Faust of remesh like the code you provided, I can't get the same accuracy as the GT you provided. The accuracy is about 20% lower, and the match I get exceeds the limit. How can I solve this problem?

[bach-zouk]

Hello,

Thank you for reporting this. 20 % drop you refer to, does it correspond to the map obtained by surfmnet or by surfmnet +pmf(refinement)?

Also, what do you mean by match exceeding the limit?

[HJ-Xu]

Hello,

I am very happy to receive your reply. I followed the code you provided to download and ran the code here. The network parameters in the test provided in your code are trained on Faust_original data, and good results can be obtained.

But I used the code you provided to train the network with the Faust_remesh data set 00-79, and then use the Faust_remesh data set 80-99 to test the network, but I still can't get the result of Supervised Methods on FAUST Remesh as shown in Figure 3 in your paper.

I used the method mentioned in your letter to get the result, and during the processing, the index will be out of bounds. For example, the remesh data A has 4999 vertices, and the matches obtained by using your test_DFMnet.py file may exist. The index value is 5000.
What should I do now? Or can I get the model parameter file that you trained on the Faust_remesh data set 00-79?

Thank you again for your reply, and sincerely ask for your answers.

[bach-zouk]

Hi,

this should not be the case. At the moment, I am a bit busy with iccv deadline. Can you perhaps try this repo in pytorch ,also written by a student in our lab? https://github.com/pvnieo/SURFMNet-pytorch
I will come back to your issues about this repo in a week. Sorry

[HJ-Xu]

hello，

I ran the code according to your recommended version of pytorch, but I still encounter the same problem.

Looking forward to your answers again when you are finished.

[bach-zouk]

Hi,

You mean accuracy issue or index out of bound issue?

[HJ-Xu]

Hi，

Thanks for your reply, I do have these problems at the moment.

But it is worth noting that these problems only occur when using code to train Faust_remesh. When I train Faust_origin and download FAUST_remeshed_matches on the https://drive.google.com/open?id=1qvqtJz-_zvMxC0ZMuFGbtlKxc9Py3Ggg you provided, the accuracy obtained after processing is the result shown in your paper, which makes me very confused.

Perhaps if the model.ckpt file of your Faust_remesh training at that time is still there, can you send a copy to the email address [email protected]? I want to use your trained model file to test again, thank you very much!