Global and Local Knowledge-Aware Attention Network for Action Recognition.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Convolutional Neural Networks (CNNs) are effective for video action recognition.
• Traditional methods often lack attention mechanisms to focus on relevant video frame parts.

Purpose of the Study:

• To propose a novel global and local knowledge-aware attention network for improved action recognition.
• To address the limitations of traditional algorithms by incorporating attention mechanisms.

Main Methods:

• A three-stream architecture combining two attention streams (statistic-based attention (SA) and learning-based attention (LA)) and a global pooling (GP) stream.
• Fusion layers to combine global and local information, producing composite features.
• Global-attention (GA) regularization to guide attention streams and softmax layer fusion for final predictions.

Main Results:

• The proposed network learns efficient video-level features spatially and temporally.
• Achieved superior performance compared to state-of-the-art methods on Kinetics, HMDB51, and UCF101 benchmarks.

Conclusions:

• The novel attention network effectively leverages global and local information for enhanced video action recognition.
• The proposed method demonstrates significant improvements on challenging action recognition datasets.