search

SizheAn / PanoHead

Code Repository for CVPR 2023 Paper "PanoHead: Geometry-Aware 3D Full-Head Synthesis in 360 degree"
Creative Commons Zero v1.0 Universal
1.92k stars 238 forks source link

Bad cases detection? #21

Closed Kizuna-AII closed 1 year ago

Kizuna-AII commented 1 year ago

Hi, thanks to your sharing of code and pretrained model! Now I am trying to generate lots of human face pictures with PanoHead, and sometimes I get some bad generation pictures, which you have mentioned in the limitation part in CVPR supplementary material. I am wondering is it possible to detect and filter bad generation results in an automatic way?

Model checkpoint: easy-khair-180-gpc0.8-trans10-025000.pkl

Thanks!

Kizuna-AII commented 1 year ago

Now I simply trained a binary classification network to detect bad cases, and it seems to work well.

Zvyozdo4ka commented 5 months ago

@Kizuna-AII HOw did you train your network? Can you share some of your outputs, which your consider as good results?

Kizuna-AII commented 5 months ago

@Kizuna-AII HOw did you train your network? Can you share some of your outputs, which your consider as good results?

Sorry that I am not able to get my previous results at this time. I just utilize a pre-trained VGG-19 model and modify the last classification layer to make it output "good" or "bad". Then I manually labelled about 1000 pictures generated by PanoHead with "good" or "bad" annotation, and fine-tune the modified VGG-19 model with these data. About the "good results" standard, you can just make a "bad" label if the picture has ANY confusing or bad-quality parts.

Zvyozdo4ka commented 5 months ago

@Kizuna-AII Thank you for explanation! I did not that PanoHead generates pictures. Are these pictures of resulting 3d model without or with texture?

Kizuna-AII commented 5 months ago

gen_samples.py will generate a picture with a given camera pose, which has texture and background.