[Paper] [Supplementary video]

This repository contains the code to train and evaluate the work presented in the article Domain and View-point Agnostic Hand Action Recognition.

@inproceedings{sabater2021domain,
  title={Domain and View-point Agnostic Hand Action Recognition},
  author={Sabater, Alberto and Alonso, I{\~n}igo and Montesano, Luis and Murillo, Ana C},
  booktitle={2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},
  year={2021},
}

Download the desired models used in the paper and store them under ./pretrained_models/.

• Last TCN descriptor
• Summarization (Best)

• F-PHAB models
• SHREC17 models

The present project uses skeleton representations based on the 20-joints that SHREC17 and F-PHAB have in common.

In F-PHAB Dataset: Wrist, TPIP, TDIP, TTIP, IMCP, IPIP, IDIP, ITIP, MMCP, MPIP, MDIP, MTIP, RMCP, RPIP, RDIP, RTIP, PMCP, PPIP, PDIP, PTIP

In SHREC-17 Dataset: Wrist, thumb_first_joint, thumb_second_joint, thumb_tip, index_base, index_first_joint, index_second_joint, index_tip, middle_base, middle_first_joint, middle_second_joint, middle_tip, ring_base, ring_first_joint, ring_second_joint, ring_tip, pinky_base, pinky_first_joint, pinky_second_joint, pinky_tip

The 7-joints minimal skeleton representation proposed in the paper uses the skeleton joints indexed by 0,8,3,7,11,15,19, which stands for Wrist, middle_base, thumb_tip, index_tip, middle_tip, ring_tip, pinky_tip.

Original skeletons files can be transformed to the 20-joints format with the scripts for SHREC17 and F-PHAB, and stored under ./datasets/. Minimal 7-joints format is later obtained with the DataGenerator.

F-PHAB data splits used for the evaluation are located under ./dataset_scripts/F_PHAB/paper_tables_annotations/.

Project tested with the following dependencies:

• python 3.6
• tensorflow 2.3.0
• Keras 2.3.1
• keras-tcn 3.1.0
• scikit-learn 0.22.2
• scipy 1.4.1
• pandas 1.0.3

Execute the file action_recognition_evaluation.py to perform the cross-domain evaluation reported in the paper. The script loads skeleton action sequences, augments them, generates sequence embeddings and performs the KNN classification. Final results show the accuracy calculated both with and without motion reference set augmentation. Use the following flags to evaluate different datasets. --eval_fphab, --eval_msra

To reproduce the results given in the Table III from the paper, download the cross-domain models and execute the following commands:

python action_recognition_evaluation.py --path_model ./pretrained_models/xdom_last_descriptor --loss_name mixknn_train5_val1 --eval_fphab

python action_recognition_evaluation.py --path_model ./pretrained_models/xdom_summarization --loss_name mixknn_best --eval_fphab --eval_msra

Note that, since random operations are involved in the evaluation, final results can slightly differ from the results reported in the paper.