search

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
1.75k stars 342 forks source link

Using Real Images for inference #95

Closed tejokrishna closed 3 years ago

tejokrishna commented 3 years ago

Hi Shunsuke,

 Thanks for sharing the code and the weights. I have a quick question, I am trying to use real world pictures which have perspective projection. Using the intrinsics, I saw that we can set the calib tensor. Is there an example that I can take a look at? Using the projection mode set to perspective is yielding bad results. Thanks in anticipation.
shunsukesaito commented 3 years ago

The released model is trained with weak-perspective projection and I don't expect it to perform well when changing it to perspective during inference. One thing you could try is to approximate perspective by weak perspective projection for PIFu inference and convert it back the original space after 3D reconstruction. Check out the following code for reference.

https://github.com/nkolot/SPIN/blob/5c796852ca7ca7373e104e8489aa5864323fbf84/train/trainer.py#L162

tejokrishna commented 3 years ago

Thank you Shusuke, that link is quite helpful.

ChanYeongShin commented 3 years ago

@tejokrishna Did you implement PIFu by using real images for inference? If you've done, can you share the code..? Thank you.