ShichenLiu / SoftRas

Project page of paper "Soft Rasterizer: A Differentiable Renderer for Image-based 3D Reasoning"

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Qualitative Results

chicleee opened this issue · comments

commented

Hi, thanks for your sharing.

I successfully trained and tested the model 250000 iter follow the usages, but the results of single-view mesh reconstruction is lower than the best socre(full: 64.64) mentioned in the paper. It is "Mean IoU: 61.838 for all classes".

Did I miss any details or Did I need to adjust some parameters?

Thanks!

The 64.6 results is trained with our own rendered shading images. You should be able to obtain mIoU over 62 with the provided data.