Enhancing Human Action Recognition with GAN-based Data Augmentation (EHAR-GAN)

Code for our 2024 paper "Enhancing human action recognition with GAN-based data augmentation," by Prasanna Reddy Pulakurthi, Celso M. de Melo, Raghuveer Rao, and Majid Rabbani. [PDF] [Dataset]

Keywords: Human Action Recognition (HAR), Generative Adversarial Network(GAN), Deep Neural Network (DNN), Synthetic Data, Data Augmentation.

Overview

EHAR-GAN proposes a GAN-based framework for enhancing human action recognition (HAR) by generating synthetic gesture videos that vary both motion and appearance.
By augmenting a small-sized real dataset with targeted motion transfer and style variation, we significantly improve HAR performance without requiring complex motion capture setups.

Datasets

• RoCoG-v2 - Original RoCoG-v2 Dataset
• Syn-RoCoG-v2 - Our Generated Synthetic Dataset

Original Video	Motion Transfer to S01	Motion Transfer to S02	Motion Transfer to S03
Motion Transfer to S05	Motion Transfer to S07	Motion Transfer to S08	Motion Transfer to S10

Results

• Experiment Results

Citation

Please consider citing our paper in your publications if it helps your research. The following is a BibTeX reference.

@inproceedings{pulakurthi2024enhancing,
  title={Enhancing human action recognition with GAN-based data augmentation},
  author={Pulakurthi, Prasanna Reddy and De Melo, Celso M and Rao, Raghuveer and Rabbani, Majid},
  booktitle={Synthetic Data for Artificial Intelligence and Machine Learning: Tools, Techniques, and Applications II},
  volume={13035},
  pages={194--204},
  year={2024},
  organization={SPIE}
}