This is the official implementation of the WACV 2025 Paper Utilizing Uncertainty in 2D Pose Detectors for Probabilistic 3D Human Mesh Recovery by Tom Wehrbein, Marco Rudolph, Bodo Rosenhahn and Bastian Wandt.
We recommend creating a clean conda environment. You can do this as follows:
conda env create -f environment.yml
A more cross-platform-friendly installation alternative could be to manually follow the commands in install_env.txt.
After the installation is complete, you can activate the conda environment by running:
conda activate humr
- SMPL: Please download version 1.1.0, extract it to
data/body_modelsand rename the model files to SMPL_NEUTRAL/FEMALE/MALE.pkl. - SMPL-X: Please download version 1.1 and extract it to
data/body_models/smplx. - SMPL Utils: Download utils.zip and extract it to
data/utils. - Pretrained Weights: Finally, place all pretrained checkpoints in
data/ckpt.
Additionally for training and evaluation, please follow the dataset preparation guide here.
We provide a few examples in data/examples. Running the following demo will detect each person in each image and then run HUMR for 3D reconstruction. The resulting rendering will be saved to logs/
python demo.py --cfg configs/default_config.yaml --ckpt data/ckpt/humr_best.ckpt
Training can be started with:
python train.py --cfg configs/default_config.yaml --name expname
To evaluate the provided trained model:
python train.py --cfg configs/default_config.yaml --ckpt data/ckpt/humr_best.ckpt --test
We benefit from many great resources including but not limited to: SPIN, CLIFF, ViTPose, HRNet, BEDLAM, ProHMR, HuManiFlow
If you find the model and code useful, please consider citing the following paper:
@InProceedings{wehrbein25humr,
author = {Wehrbein, Tom and Rudolph, Marco and Rosenhahn, Bodo and Wandt, Bastian},
title = {Utilizing Uncertainty in 2D Pose Detectors for Probabilistic 3D Human Mesh Recovery},
booktitle = {Winter Conference on Applications of Computer Vision (WACV)},
year = {2025},
}