Taein Kwon, Bugra Tekin, Sigyu Tang, and Marc Pollefeys

This code is for Context-Aware Sequence Alignment using 4D Skeletal Augmentation, CVPR 2022 (oral). You can see more details about more paper in our project page. Note that we referred LoFTR to implement our framework.

To setup the env,

git clone https://github.com/taeinkwon/casa_clean.git
cd CASA
conda env create -f env.yml
conda activate CASA

.
├── bodymocap
├── extra_data
│   └── body_module
│         └── J_regressor_extra_smplx.npy
├── human_body_prior
│   └── ...
├── manopth
│   ├── __init__.py
│   ├── arguitls.py
│   └── ...
├── smpl
│   └── models
│         └── SMPLX_NEUTRAL.pkl
├── mano 
│   ├── models
│   │     ├── MANO_LEFT.pkl
│   │     └── MANO_RIGHT.pkl
│   ├── websuers
│   ├── __init__.py
│   └── LICENSE.txt
├── npyrecords
├── sripts
├── src
└── ...

In this repository, we use the MANO model from MPI and some part of Yana's code for hand pose alignment.

• Clone manopth git clone https://github.com/hassony2/manopth.git and copy manopth and mano folder (inside) into the CASA folder.
• Go to the mano website and download models and code and put them in CASA/mano/models.
• In smpl_handpca_wrapper_HAND_only.py, please change following lines to run in python3. L23:import CPickle as pickle -> import pickle, L30: dd = pickle.load(open(fname_or_dick)) -> dd = pickle.load(open(fname_or_dict, 'rb'), encoding='latin1'), L144: print 'FINTO' -> print('FINTO').

We used Vposer to augment body pose.

• In the Vposer repository, clone it git clone https://github.com/nghorbani/human_body_prior.git
• Copy the human_body_prior folder into the CASA folder.
• Go into the human_body_prior folder and run the setup.py

  cd human_body_prior
  python setup.py develop
  

We use the SMPL model for body pose alignment.

• Download SMPL-X ver 1.1. and VPoser V2.0.
• Put the SMPLX_NUETRAL.pkl into the CASA/smpl/models folder.
• Copy VPoser files in CASA/human_body_prior/support_data/downlaods/vposer_v2_05/
• In order to use the joints from FrankMocap, you need to additionally clone FrankMocap git clone https://github.com/facebookresearch/frankmocap.git and put the bodymocap fodler into CASA fodler.
• After then, run sh download_extra_data.sh to get the J_regressor_extra_smplx.npy file.

You can download npy files here. In the npy files, normalized joints and labels are included. In order to get the original data, you should go to each dataset websites and download the datasets there.

We selected pouring milk sequences and manually divided into train and test set with the new labels we set. Please go to the H2O project page and download the dataset there.

We estimated 3D joints using FrankMocap for the Penn Action dataset. Penn Action has 13 different actions: baseball_pitch, baseball_swing, bench_press, bowling, clean_and_jerk, golf_swing, jumping_jacks, pushups, pullups, situp, squats, tennis_forehand, tennis_serve.

We downloaded and used the 3D joints from triangulation of 2D poses in the IkeaASM dataset.

To train the Penn Action dataset,

sh scripts/train/pennaction_train.sh ${dataset_name}

For example,

sh scripts/train/pennaction_train.sh tennis_serve

We also provide pre-trained models. To evalate the pre-trained model

sh scripts/eval/pennaction_eval.sh ${dataset_name} ${eval_model_path}

For example,

sh scripts/eval/pennaction_eval.sh tennis_serve logs/tennis_serve/CASA=64/version_0/checkpoints/last.ckpt

Note that our code follows the Apache License 2.0. However, external libraries follows their own licenses.