PyTorch implementation of CloudWalk's recent paper DenseBody.

Note: For most recent updates, please check out the dev branch.

Update on 20190613 A toy dataset has been released to facilitate the reproduction of this project. checkout PREPS.md for details.

Update on 20190826 A pre-trained model (Encoder/Decoder) has been released to facilitate the reproduction of this project.

Here is the reproduction result (left: input image; middle: ground truth UV position map; right: estimated UV position map)

• SMPL official UV map is now supported! Please checkout PREPS.md for details.
• Code reformating complete! Please refer to data_utils/UV_map_generator.py for more details.
• Thanks Raj Advani for providing new hand crafted UV maps!

Please follow the instructions PREPS.md to prepare your training dataset and UV maps. Then run train.sh or nohup_train.sh to begin training.

To train with your own UV map, checkout UV_MAPS.md for detailed instructions.

To explore different network architectures, checkout NETWORKS.md for detailed instructions.

• Creating ground truth UV position maps for Human36m dataset.

  • 20190329 Finish UV data processing.
  • 20190331 Align SMPL mesh with input image.
  • 20190404 Data washing: Image resize to 256*256 and 2D annotation compensation.
  • 20190411 Generate and save UV position map.
    • radvani Hand parsed new 3D UV data
    • Validity checked with minor artifacts (see results below)
    • Making UV_map generation module a separate class.
  • 20190413 Prepare ground truth UV maps for washed dataset.
  • 20190417 SMPL official UV map supported!
  • 20190613 A testing toy dataset has been released!

• Prepare baseline model training

  • 20190414 Network design, configs, trainer and dataloader
  • 20190414 Baseline complete with first-hand results. Something issue still needs to be addressed.
  • 20190420 Testing with different UV maps.

Lingbo Yang(Lotayou): The owner and maintainer of this repo.

Raj Advani(radvani): Provide several hand-crafted UV maps and many constructive feedbacks.

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

The network training part is inspired by BicycleGAN