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

Partial views #71

Closed faniry6 closed 3 years ago

faniry6 commented 3 years ago

Hi Shunsuke

In our training datasets, we have views where only a part of the object appear. This may cause problems since the prediction might be then wrong especially at the stage where you use

in_img = (xy[:, 0] >= -1.0) & (xy[:, 0] <= 1.0) & (xy[:, 1] >= -1.0) & (xy[:, 1] <= 1.0) in HGPIFuNet.py

A point can be inside the geometry even though it is not seen in the current view.

Thanks in advance

shunsukesaito commented 3 years ago

This is after normalizing pixel coordinates to [-1, 1]. PIFu is reasoning 3D shapes from corresponding pixel-aligned features, meaning it's not possible to make any prediction for outside image plane. The reason why the value is ranged in [-1, 1] is because this is the expected value for F.grid_sample.