search

chaitanya100100 / TailorNet

Code for our CVPR 2020 (ORAL) paper - TailorNet: Predicting Clothing in 3D as a Function of Human Pose, Shape and Garment Style.
https://virtualhumans.mpi-inf.mpg.de/tailornet/
Other
408 stars 67 forks source link

Graph CNN code for Single style-Shape model #1

Closed yuxwind closed 4 years ago

yuxwind commented 4 years ago

Great job! The paper reported an experiment of Graph CNN for the style-shape model. It is interesting that MLP outperforms GCN here. I would like to test some ideas based on your GCN network. Would you plan to release its code? It is also good if you can provide detailed network settings. Thanks a lot!

chaitanya100100 commented 4 years ago

Hi yuxwind, I am glad that you find our paper interesting. :)

Currently, we don't have any plan to release GraphCNN code. It's not a small code and it requires some preprocessing steps. Including that here would divert the focus of the code. And it's not the main focus of the paper.

Having said that, I will try to list some key aspects of the Graph CNN we used. It shouldn't take much effort to implement it.

  1. We use the formulation of [1] for graph CNN. [2] has done the graphCNN regression on SMPL body, so we followed the same architecture.
  2. Our garment template is too high-resolution for training with GraphCNN. So we downsample the template by a factor of 3. [3] might be useful for that.

GraphCNNs haven't shown promising results in tasks which require fine grained results like garment wrinkles.

[1] Thomas N. Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. In International Conference on Learning Representations (ICLR), 2017. [2] https://www.seas.upenn.edu/~nkolot/projects/cmr/ [3] https://github.com/anuragranj/coma/blob/master/lib/mesh_sampling.py