jiaxiangshang / MGCNet

Self-Supervised Monocular 3D Face Reconstruction by Occlusion-Aware Multi-view Geometry Consistency[ECCV 2020]

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

the testing code does not work

puppet101 opened this issue · comments

Hi, I used the testing code in the test_unsupervise.py, but some errors occurred. Is the testing code unfinished?
Thank you!

Yes, the test code is unfinished, I try to finish it before August 15.

Hi, @jiaxiangshang
Thank you for finishing the testing code. But the result is a little different from which it shows in the demo.
For example, the color of this image is a little strange, and the shape is also not as perfect as it is in your demo.

image04275_mulPoses

Really thanks for your quick feed back.
The color is sure of a problem. I will first fix this and tell you.
For the shape, I will test more images and see if this is because of the new model that I trained with this release version code.

Thanks a lot for your excellent work!

The color bug now fixed.
For the shape rendered on images, the overlay result alignment mainly depends on the camera pose rather than the face shape, the face shape can be visualized in meshlab that is still good.
For the camera pose, I will check the model I uploaded, I will update tomorrow.

Thanks a lot for your excellent work!

Hi, puppet101
I have a test on other images, it shows consistency results, so the model is ok.
This sample not as perfect as our paper may because of retraining and a little changes of the training dataset.
Welcome give further feedback and suggestion, I will keep work on face reconstruction.

Great! Thank you again for your help!