Towards Accurate Cross Domain In-Bed Human Pose Estimation (Accepted at ICASSP 2022)

Citation

If you find our work or this repository useful, please consider giving a star ⭐ and citation.

@article{CD_HPE,
      title={Towards Accurate Cross-Domain In-Bed Human Pose Estimation}, 
      author={Mohamed Afham and Udith Haputhanthri and Jathurshan Pradeepkumar and Mithunjha Anandakumar and Ashwin De Silva and Chamira Edussooriya},
      year={2021},
      journal={arXiv preprint arXiv:2110.03578},
}

Pose Definition

We follow the Leeds Sports Pose Dataset pose definition with 14 joints labeling, namely,

Right ankle Right knee Right hip Left hip Left knee Left ankle Right wrist Right elbow Right shoulder Left shoulder Left elbow Left wrist Neck Head top

The lable matrix joints_gt_<modality>.mat has the format <x,y,if_occluded> x n_joints x n_subjects

Dataset

Create a folder named data/SLP_VIPCup.

mkdir data
mkdir data/SLP_VIPCup

Download dataset from codalab. Unzip the data inside data/SLP_VIPCup/ by running:

source download_data.sh

The directory should look like

CD_HPE/ 
    data/
        SLP_VIPCup/ 
            train/ 
            test1/ 
            val/

Run the following to create the required .json files for the dataset

cd filelists
source create_filelist.sh
cd ..

Train

For Learning Stage 01 (Standard Supervision)

python train_supervised.py --adam --use_target_weight --model stacked_hg --print_freq 50 --batch_size 3 --wandb_run Name/of/wandb/run

For Learning Stage 02 (Knowledge Distillation)

python train_distillation.py --adam --best_path path/to/best/model --model stacked_hg --print_freq 50 --batch_size 4 --lr 1e-4 --wandb_run Name/of/wandb/run

Kindly refer to for reproducing our results.