LIUkhan / DeePSD

This repository contains the necessary code to run the model described in:
https://arxiv.org/abs/2009.02715

DATA

The dataset used on this work and this repository is CLOTH3D, with associated paper.
Path to data has to be specified at 'values.py'. Note that it also asks for the path to preprocessings, described below.

PREPROCESSING

In order to optimize data pipeline, we preprocess template outfits. The code to train the model assumes the preprocessing is done. To perform this preprocessing, check the scripts at 'DeePSD/Preprocessing/'.

1. outfit_verts.py It creates 'txt' files for each sample. It relies on this for garment-to-outfit and outfit-to-garment conversions.
2. rest.py Rest garments are stored in OBJ files, which are encoded in ASCII. To increase efficiency, we convert this into binary format.
3. faces2edges.py Precomputes list of edges as [v0, v1] as int16 in binary format.
4. laplacians.py Precomputes laplacian matrices for each outfit.
5. weights_prior.py Precomputes blend weights labels by proximity to body in rest pose. Used for guiding learning in the first epoch.

TRAIN

Once all preprocessings have been completed. Just run 'train.py' and 'train_chi.py'.

SMPL

We removed SMPL models in PKL format due to their size. The code will expect those as '/DeePSD/Model/smpl/model_f.pkl' and '/DeePSD/Model/smpl/model_m.pkl'.