NVlabs / nvdiffrec

Official code for the CVPR 2022 (oral) paper "Extracting Triangular 3D Models, Materials, and Lighting From Images".
Other
2.13k stars 223 forks source link

Rendering quality depends too much on foreground masks quality #86

Closed iraj465 closed 1 year ago

iraj465 commented 1 year ago

Hi, Great work on the paper! Really good stuff. I was able to get some mid level quality results on my own datasets using rembg alpha channels. I was looking for manual NERD based masking to get better results. Not upto much good. Can you please provide any code which you have used in your paper with Detectron2 etc for generating masks? I see your results to be really good, might be that the masks' quality played a good role. Any leads is appreciated

jmunkberg commented 1 year ago

Thanks @iraj465 ,

For the mask tests in the supplemental, we used two versions of Detectron2, namely PointRend and Mask R-CNN. We used the code from the Detectron 2 github repo: https://github.com/facebookresearch/detectron2

This service also worked quite well in some tests: https://www.remove.bg/

Good masks definitely help our reconstructions.

iraj465 commented 1 year ago

Gotcha, thanks for this!