AR3D: Attention Residual 3D Network for Human Action Recognition.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• 2D CNNs independently process temporal and spatial features, limiting human action recognition.
• 3D CNNs capture spatio-temporal features but suffer from high parameter counts and training difficulties.

Purpose of the Study:

• To address limitations of existing deep learning models for video-based human action recognition.
• To propose novel 3D CNN architectures that are efficient and effective.

Main Methods:

• Developed a shallow feature extraction module to enhance temporal feature extraction in 3D CNNs.
• Integrated a 3D spatio-temporal attention mechanism to improve global feature extraction.
• Proposed Residual 3D Network (R3D) and Attention Residual 3D Network (AR3D) models, including fusion strategies (AR3D_V1, AR3D_V2).

Main Results:

• The improved 3D residual structure reduces parameters and strengthens temporal feature extraction.
• The 3D spatio-temporal attention module enhances global feature extraction for human actions.
• Experimental results demonstrate performance improvements with fused R3D and AR3D structures compared to single components.

Conclusions:

• The proposed R3D and AR3D models offer improved performance in human action recognition.
• Combining residual structures and attention mechanisms in 3D CNNs is effective for enhancing spatio-temporal feature learning.
• The novel architectures provide a more efficient and robust solution for video-based action recognition tasks.