search

lijiaman / egoego_release

Official Implementation of the Paper: Ego-Body Pose Estimation via Ego-Head Pose Estimation (CVPR 2023)
MIT License
124 stars 11 forks source link

readme

Ego-Body Pose Estimation via Ego-Head Pose Estimation (CVPR 2023)

This is the official implementation for the CVPR 2023 paper. For more information, please check the project webpage.

EgoEgo Teaser

Environment Setup

Note: This code was developed on Ubuntu 20.04 with Python 3.8, CUDA 11.3 and PyTorch 1.11.0.

Clone the repo.

git clone https://github.com/lijiaman/egoego.git
cd egoego/

Create a virtual environment using Conda and activate the environment. 

conda create -n egoego_env python=3.8
conda activate egoego_env

Install PyTorch. 

conda install pytorch==1.11.0 torchvision==0.12.0 torchaudio==0.11.0 cudatoolkit=11.3 -c pytorch

Install PyTorch3D. 

conda install -c fvcore -c iopath -c conda-forge fvcore iopath
conda install -c bottler nvidiacub
pip install --no-index --no-cache-dir pytorch3d -f https://dl.fbaipublicfiles.com/pytorch3d/packaging/wheels/py38_cu113_pyt1110/download.html

Install human_body_prior. 

git clone https://github.com/nghorbani/human_body_prior.git
pip install tqdm dotmap PyYAML omegaconf loguru
cd human_body_prior/
python setup.py develop

Install mujoco. 

wget https://github.com/deepmind/mujoco/releases/download/2.1.0/mujoco210-linux-x86_64.tar.gz
tar -xzf mujoco210-linux-x86_64.tar.gz
mkdir ~/.mujoco
mv mujoco210 ~/.mujoco/
export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:~/.mujoco/mujoco210/bin

Install other dependencies. 

pip install evo --upgrade --no-binary evo
pip install -r requirements.txt

Quick Start

First, download pretrained models and put pretrained_models/ to the root folder.

If you would like to generate visualizations, please download Blender first. And put blender path to blender_path. Replace the blender_path in line 45 of egoego/vis/blender_vis_mesh_motion.py.

Please download SMPL-H (select the extended SMPL+H model) and put the model to smpl_models/smplh_amass/. If you have a different folder path for SMPL-H model, please modify the path in line 13 of egoego/data/amass_diffusion_dataset.py.

Then run EgoEgo pipeline on the testing data. This will generate corresponding visualization results in folder test_data_res/. To disable visualizations, please remove --gen_vis. 

sh scripts/test_egoego_pipeline.sh

Datasets

If you would like to train each module of our pipeline and evaluate, please prepare the following datasets.

AMASS

Please download AMASS data following the instructions on the website.

We used SMPL-H data for this project. Please put AMASS data to a folder data/amass.

cd utils/data_utils 
python process_amass_dataset.py 
python convert_amass_to_qpos.py

Please replace the data path with your desired path in the file before running.

ARES

ARES dataset relies on AMASS. Please processing AMASS data following above instruction first before processing ARES.

First, download egocentric videos and data that contains AMASS sequence information using this link.

Please put the data to a folder data/ares. Uncompress the files in ares_ego_videos.

To extract corresponding motion sequences from AMASS for each egocentric video, please run the following command. This will copy each egocentric video's corresponding motion to data/ares/ares_ego_videos/scene_name/seq_name/ori_motion_seq.npz.

cd utils/data_utils 
python extract_amass_motion_for_ares.py

To prepare data used during training, run this command to convert all the motion sequences into a single data file. Please check data path before you proceed. 

python process_ares_dataset.py

To prepare data used during evaluation, run this command to convert data format to be consistent with kinpoly. Please check data path before you proceed. 

python convert_ares_to_qpos.py

Kinpoly

Please download Kinpoly data following the instructions in this repo.

Put MoCap dataset to folder data/kinpoly-mocap/.

Put RealWorld dataset to folder data/kinpoly-realworld/.

GIMO

Please download GIMO data following the instructions in this repo.

To process GIMO, please do the following. 

cd utils/data_utils/gimo_utils
python segment_seq_images.py

Then extract pose parmaters from Vposer.

python extract_pose_params.py

Process GIMO to be consistent with AMASS processed data.

cd utils/data_utils
python process_gimo_data.py

Process GIMO data to the format used for training and evaluation.

cd utils/data_utils
python convert_gimo_to_qpos.py

Others

We used DROID-SLAM to extract camera poses. We also provided results of DROID-SLAM for ARES, Kinpoly-MoCap, GIMO here. Please find the results for each dataset and put them into desired path data/ares/droid_slam_res/, data/gimo/droid_slam_res/, data/kinpoly-mocap/droid_slam_res, data/kinpoly-realworld/droid_slam_res/.

We used RAFT to extract optical flow, we provided optical flow features extracted using a pre-trained ResNet here. Please find the results for each dataset and put them into desired path data/ares/raft_of_feats, data/gimo/raft_of_feats, data/kinpoly/fpv_of_feats.

Evaluation

Evaluate the conditional diffusion model at stage 2 on AMASS testing split. This part only relies on AMASS data. To generate visualizations, please add --gen_vis. 

sh scripts/eval_stage2.sh

Evaluate the whole EgoEgo pipeline on ARES, GIMO, Kinpoly-MoCap. Please use --test_on_ares, --test_on_gimo, and --eval_on_kinpoly_mocap respectively. To generate visualizations, please add --gen_vis. Note that before proceeding, please download the sequence names that are not included in quantitative evaluation as DROID-SLAM failed in these sequences. Our approach relies on a reasonable SLAM result. Put the folder to data/failed_seq_names/. 

sh scripts/eval_egoego_pipeline_on_ares.sh
sh scripts/eval_egoego_pipeline_on_gimo.sh
sh scripts/eval_egoego_pipeline_on_kinpoly.sh

In utils/blender_utils folder, we provided multiple .blend files as reference for you to generate visualizations. You can try different .blend or customize your own visualization by modifying .blend using Blender.

Training

For all the script used in training, please modify --entity to your username on wandb to monitor the training loss.

Training Conditional Diffusion Model for Full-Body Generation from Head Pose

Train conditional diffusion on AMASS.

sh scripts/train_full_body_cond_diffusion.sh

Training GravityNet

sh scripts/train_gravitynet.sh

Training HeadNet

Train HeadNet on ARES.

sh scripts/train_headnet_on_ares.sh

Train HeadNet on GIMO. 

sh scripts/train_headnet_on_gimo.sh

Train HeadNet on Kinpoly-Realworld. 

sh scripts/train_headnet_on_kinpoly.sh

Citation

@inproceedings{li2023ego,
  title={Ego-Body Pose Estimation via Ego-Head Pose Estimation},
  author={Li, Jiaman and Liu, Karen and Wu, Jiajun},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={17142--17151},
  year={2023}
}

Related Repos

We adapted some code from other repos in data processing, learning, evaluation, etc. Please check these useful repos. 

https://github.com/lucidrains/denoising-diffusion-pytorch
https://github.com/davrempe/humor
https://github.com/KlabCMU/kin-poly 
https://github.com/jihoonerd/Conditional-Motion-In-Betweening 
https://github.com/lijiaman/motion_transformer