Skip to content

README.md

Latest commit

 

History

History
80 lines (66 loc) · 6.4 KB

README.md

File metadata and controls

80 lines (66 loc) · 6.4 KB
Raw

Multimodal-Action-Recognition

This repo is the official implementation for EPAM-Net: An Efficient Pose-driven Attention-guided Multimodal Network for Video Action Recognition, which is accepted by Neurocomputing 2025. Paper
Paper: https://www.sciencedirect.com/science/article/pii/S0925231225004539

Abstract

Existing multimodal-based human action recognition approaches are either computationally expensive, which limits their applicability in real-time scenarios, or fail to exploit the spatial temporal information of multiple data modalities. In this work, we present a novel and efficient pose-driven attention-guided multimodal network (EPAM-Net) for action recognition in videos. Specifically, we propose eXpand temporal Shift (X-ShiftNet) convolutional architectures for RGB and pose streams to capture spatio-temporal features from RGB videos and their skeleton sequences. The X-ShiftNet tackles the high computational cost of the 3D CNNs by integrating the Temporal Shift Module (TSM) into an efficient 2D CNN, enabling efficient spatiotemporal learning. Then skeleton features are utilized to guide the visual network stream, focusing on keyframes and their salient spatial regions using the proposed spatial–temporal attention block. Finally, the predictions of the two streams are fused for final classification.. The proposed architecture achieved comparative performance with state-of-the-art methods on NTU RGB-D 60, NTU RGB-D 120, PKU-MMD, and Toyota SmartHome datasets with up to a 72.8x reduction in FLOPs and up to a 48.6x reduction in the number of network parameters.

Architecture of EPAM-Net

Prerequisites

  • Python3
  • PyTorch
  • We provide the dependency file of our experimental environment, you can install all dependencies by creating a new anaconda virtual environment and running pip install -r requirements.txt

Download dataset

  1. NTU-RGB+D 60 dataset from https://rose1.ntu.edu.sg/dataset/actionRecognition/
  2. NTU-RGB+D 120 dataset from https://rose1.ntu.edu.sg/dataset/actionRecognition/
  3. PKU-MMD dataset from https://www.icst.pku.edu.cn/struct/Projects/PKUMMD.html
  4. Toyota SmartHome trimmed dataset from https://project.inria.fr/toyotasmarthome/

Data Preparation

For Pose estimation, we utilize a Top-Down pose estimation approach instantiated with HRNet-W32 to extract 2D poses from videos and save the coordinate triplets (x, y, score) following PoseConv work

Here, links to the pre-processed skeleton annotations, you can directly download them and use them for training & testing. For NTU 60 and 120 dataset, you can use the script file data_preparation/prepare_ntu_dataset_annotations.py to split such pose annotations (ntu120_2d.pkl) into X-Sub, X-Set for NTU120 and similarly for NTU60.

For RGB videos, We crop the RGB frames with the global box, which envelops all persons in the video, and resize them to a resolution of 224 x 224 using script data_preparation/crop_RGBFrames_PKU.py (PKU dataset is an example). The cropped RGB frames are used as input for the RGB stream.

For PKU, since it is an untrimmed dataset, we extract the actions clips from each video using the script data_preparation/pku_dataset_videos_split.py.

Testing Pretrained Models

You may download the trained models reported in the paper via GoogleDrive and put them in folder pretrained_EPAM_models.

Evaluation

You can use the following command to test a multimodal EPAM-Net on NTU 60, NTU 120, PKU and Toyota Smarthome datasets.

python inference/multimodal_evaluation.py --dataset 'toyota' --evaluation_protocol 'xsub'

Result

We report the mean accuracy (%) for Toyota-Smarthome dataset and Top-1 accuracy (%) for other datasets using 1-clip per video.

Method NTU-60 X-Sub NTU-60 X-View NTU-120 X-Sub NTU-120 X-Set PKU X-Sub PKU X-View Toyota SH X-Sub Toyota SH X-View2
EPAM-Net 96.1% 99.0% 92.4% 94.3% 96.2% 98.4% 71.7% 67.8%

Train

To train a new EPAM-Net, you need to train submodels for two inputs: skeleton joint, and RGB video, or you can download the pretrained submodels via GoogleDrive. After that, you can use the following command to train a unimodal model.

python train/ddp_train_MMActionRecognition_model_pytorch.py --dataset 'ntu120' --evaluation_protocol 'xsub' --skeleton_train_pkl '/datasets_annot/ntu120_xsub_train.pkl' --skeleton_test_pkl '/datasets_annot/ntu120_xsub_val.pkl' --PretrainedPose 'ntu120_xsub_x3dTShiftPose_SE_best_top1_acc_epoch_235.pth' --PretrainedRGB 'ntu120_xsub_x3dTShiftD_I3Dhead_RGB_epoch_205.pth'

Comparison between Multimodal HAR Approaches

Acknowledgements

This repo is based on MMAction2.

Citation

Please cite this work if you find it useful:

@article{abdelkawy2025epam,
  title={EPAM-Net: An efficient pose-driven attention-guided multimodal network for video action recognition},
  author={Abdelkawy, Ahmed and Ali, Asem and Farag, Aly},
  journal={Neurocomputing},
  pages={129781},
  year={2025},
  publisher={Elsevier}
}