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 343 forks source link

Why Use render img over original image? #114

Open LazerLikeFocus opened 2 years ago

LazerLikeFocus commented 2 years ago

I have see that in paper you are trying to convert img to 3d model. So the img should be real img (camera clicked)

SO why is Pifu using a rendered img for training?

https://github.com/shunsukesaito/PIFu/blob/02a6d8643d3267d06cb4a510b30a59e7e07255de/apps/render_data.py#L193 https://github.com/shunsukesaito/PIFu/blob/02a6d8643d3267d06cb4a510b30a59e7e07255de/apps/render_data.py#L194 https://github.com/shunsukesaito/PIFu/blob/02a6d8643d3267d06cb4a510b30a59e7e07255de/apps/render_data.py#L243

Why dont we just train using the original jpg img? Can't we just resize it ?

kampelmuehler commented 2 years ago

Well. PIFu is trained on textured meshes, so you need to render them somehow. If of course you do have the raw data used for generating those you could train on the actual images.