xjtueducation / densebody_pytorch

PyTorch implementation of CloudWalk's recent work DenseBody

densebody_pytorch

PyTorch implementation of CloudWalk's recent paper DenseBody

Critical Warning

Anyone help with UV data correction will be deeply appreciated!

SMPL UV data downloaded from official website is a total mess up. Here's the result.

Prerequisites

Ubuntu 18.04
CUDA 9.0
Python 3.6
PyTorch 1.0.0
chumpy (For converting SMPL model to basic numpy arrays)
spacepy, h5py (For processing Human36m cdf annotations)

(Optional) Install torch-batched-svd for speedup (Only tested under Ubuntu system).

TODO List

• Creating ground truth UV position maps for Human36m dataset.
  • 20190329 Finish UV data processing.
  • 20190331 Align SMPL mesh with input image.
  • Testing Generate and save UV position map.
    • Proceeding Checking validity through resampling and mesh reconstruction...
    • Making UV_map generation module a separate class.
  • Data washing: Image resize to 256*256 and 2D annotation compensation.
  • Data Preparation.
• Finish baseline model training
  • Testing with several new loss functions.
  • Testing with different networks.
• Report 3D reconstruction results.
  • Setup evaluation protocal and MPJPE-PA metrics.

Current Progress

Finish UV texture map processing. Here's the result:

Align SMPL meshes with input images. Here are some results:

Citation

Please consider citing the following paper if you find this project useful.

DenseBody: Directly Regressing Dense 3D Human Pose and Shape From a Single Color Image

Disclaimer

Please note that this is an unofficial implementation free for non-commercial usage only. For commercial cooperation please contact the original authors.