shunsukesaito / PIFu

This repository contains the code for the paper "PIFu: Pixel-Aligned Implicit Function for High-Resolution Clothed Human Digitization"
https://shunsukesaito.github.io/PIFu/
Other
1.76k stars 340 forks source link

Add 2 phase Generate Option. #6

Closed kotauchisunsun closed 4 years ago

kotauchisunsun commented 4 years ago

I want to generate the model with high quality texture. So, I add the feature to generate the sdf to file and generate the vertex color from file.

--only_sdf file The program make the sdf data and save the file.

--load_sdf file The program skip to generate the sdf and load from the file.

I made sample script.

./scripts/eval_two_phase.sh sample_images/  VOXEL_RESOLUTION VOXEL_LOAD_SIZE TEX_LOAD_SIZE

In my experience, I get the good model to set the VOXEL_LOAD_SIZE=512. But, I get the low quality texture to set the the TEX_LOAD_SIZE=512. I have to set the TEX_LOAD_SIZE=1024,2048... to get the high quality texture.

shunsukesaito commented 4 years ago

Hi, thanks for contributing to the PIFu repo! I read through the code and unfortunately this modification is not using the pretrained model properly. Thus, I cannot merge it.

Both netG and netC are trained with 512 x 512 images and behave reasonably only with these resolutions. Although increasing input resolution for netC may improve the frontside texture as it's more or less copy pasting input image, the backside gets messed up as you can see in the attached screenshot.

If you want to obtain more detailed texture for the front side, I would recommend simply projecting the input image onto the mesh and combining it with PIFu prediction at 512 x 512 for non-visible surface.

image

kotauchisunsun commented 4 years ago

@shunsukesaito Thank you for review. I close this pull request.