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
1.77k stars 340 forks source link

Training with Real Images Captured by Camera #83

Closed 475651582 closed 3 years ago

475651582 commented 3 years ago

Hi, thanks for your code! It's really a great work! I am training PIFu with images that captured by camera and I change the projection mode to 'perspective'. However, I found the speed of convergence is much slower than orthogonal mode. Does it make sense? Thank you in advance!

shunsukesaito commented 3 years ago

Generally speaking, depth normalization is the key to stably train PIFu. For orthogonal, this can be easily achieved as depth center can be arbitrarily set. However, for perspective, you need to be careful about depth range each data sample takes. If they are not well normalized, convergence can be extremely slow.