shunsukesaito / SCANimate

This repository contains the code for the paper "SCANimate: Weakly Supervised Learning of Skinned Clothed Avatar Networks"
https://scanimate.is.tue.mpg.de/
Other
274 stars 31 forks source link

question: how to get high frequency/resolution details in the generated animation mesh #4

Closed codesavory closed 2 years ago

codesavory commented 3 years ago

I wanted to know if there is a reason why the generated animated mesh is flattened/smooth with only the low frequency details even though the input mesh has high frequency features. Is there a parameter or option to get the results in high frequency fidelity?

Input Mesh - input_scene1 Generated Mesh - generated_mesh1

shunsukesaito commented 3 years ago

Looks like it's not trained long enough. Can you train longer to see if details appear?

SUNRISEFLY commented 3 years ago

@codesavory I get the similar results that lack high-level details .It seems that canonical meshs and reposed meshs in intermediate directory (./demo_result/runs) are OK. So i think the IGRSDFNET may have some trouble.To solve this , I try to train in more epoches but the question remains. r u successful now? And what's your solution?

codesavory commented 3 years ago

Hey @SUNRISEFLY , I retrained the model with longer epochs and it definitely increased the quality of the generated scanimats.

I ran this for 10k epochs which ran for than an hour on my RTX2060 Laptop, but the results are pretty sweet - 10k_epochs

1k epochs which ran for about 10minutes - 1k_epochs

JinlongYANG commented 3 years ago

Besides increasing the epochs, you can also try to reduce the learning rate for IGR. For example epochs from 1000 to 3000, learning rate from 0.004 to 0.001, I got: image