zju3dv / SA-HMR

Code for "Learning Human Mesh Recovery in 3D Scenes" CVPR 2023

Learning Human Mesh Recovery in 3D Scenes

Project Page | Paper

Learning Human Mesh Recovery in 3D Scenes
Zehong Shen, Zhi Cen, Sida Peng, Qing Shuai, Hujun Bao, Xiaowei Zhou
CVPR 2023

Setup

Environment

conda create -y -n sahmr python=3.8
conda activate sahmr
pip install -r requirements.txt
pip install -e . 

# torchsparse==1.4.0, please refer to https://github.com/mit-han-lab/torchsparse
sudo apt-get install libsparsehash-dev
pip install --upgrade git+https://github.com/mit-han-lab/torchsparse.git@v1.4.0

Weights and data

🚩 Google drive link

Model Weights

We provide the pretrained rich and prox models for evaluation under the release folder.

RICH/PROX dataset

Evaluation

1. You need to agree and follow the RICH dataset license and the PROX dataset license to use the data.

2. Here, we provide the minimal and pre-propcessed RICH/sahmr_support and PROX/quantitative/sahmr_support for reproducing the metrics in the paper. By downloading, you agree to the RICH dataset license and the PROX dataset license.

Training ✨

1. You need to submit a request to the authors from MPI and use their links for downloading the full datasets.

2. RICH: We use the JPG format image. We downsampled the image to one-forth of its original dimensions.

Link weights and data to the project folder

datasymlinks
├── RICH
│   ├── images_ds4      # see comments below
│   │   ├── train
│   │   └── val
│   ├── bodies          # included in the RICH_train.zip
│   │   ├── train
│   │   └── val
│   └── sahmr_support
│       ├── scene_info  # included in the RICH.zip
│       ├── test_split  # included in the RICH.zip
│       ├── train_split # included in the RICH_train.zip
│       └── val_split   # included in the RICH_train.zip
├── PROX                # included in the PROX.zip
└── checkpoints
    ├── release         # included in the `release`
    │   ├── sahmr_rich_e30.pth
    │   └── sahmr_prox_e30.pth
    └── metro           # see comments below
        └── metro_3dpw_state_dict.bin 

• images_ds4: Please download the train and val datasets and downsample the images to one-forth of its original dimensions.

• bodies: We provide the fitted smplh parameters for each image. We will shift to the original smplx parameters in the future.

• metro_3dpw_state_dict.bin: You only need this if you want to do training.

  Download the pretrained weights of METRO

  mkdir -p datasymlinks/checkpoints/metro
  # See https://github.com/microsoft/MeshTransformer/blob/main/LICENSE
  # See https://github.com/microsoft/MeshTransformer/blob/main/scripts/download_models.sh
  wget -nc https://datarelease.blob.core.windows.net/metro/models/metro_3dpw_state_dict.bin -O datasymlinks/checkpoints/metro/metro_3dpw_state_dict.bin

ln -s path-to-models(smpl-models) models

mkdir datasymlinks
mkdir -p datasymlinks/checkpoints
ln -s path-to-release(weights) datasymlinks/checkpoints/release

# the RICH folder should contain the original RICH dataset in the training phase,
# and the `RICH/sahmr_support` is enough for evaluation
mkdir -p datasymlinks/RICH 
ln -s path-to-rich-sahmr_support datasymlinks/RICH/sahmr_support
# for the training parts, please refer to the folder structure above

# the `PROX/quantitative/sahmr_support` is enough for evaluation
mkdir -p datasymlinks/PROX/quantitative
ln -s path-to-prox-sahmr_support datasymlinks/PROX/quantitative/sahmr_support

Usage

Evaluation

# RICH model
python tools/dump_results.py -c configs/pose/sahmr_eval/rich.yaml 
python tools/eval_results.py -c configs/pose/sahmr_eval/rich.yaml

# PROX model
python tools/dump_results.py -c configs/pose/sahmr_eval/prox.yaml
python tools/eval_results.py -c configs/pose/sahmr_eval/prox.yaml

Training

# We provide a training example on RICH dataset
python train_net.py -c configs/pose/rich/rcnet.yaml 
python train_net.py -c configs/pose/rich/sahmr.yaml 

Citation

@article{shen2023sahmr,
    title={Learning Human Mesh Recovery in 3D Scenes},
    author={Shen, Zehong and Cen, Zhi and Peng, Sida and Shuai, Qing and Bao, Hujun and Zhou, Xiaowei},
    journal={CVPR},
    year={2023}
}