Learning Semantic-Aligned Action Representation.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Action recognition relies on discriminative features, but local motion/appearance features often lack semantic alignment.
• Current methods struggle with coarse global feature pooling due to the absence of body part or object association in local features.

Purpose of the Study:

• To develop a deep convolutional neural network (DCNN) architecture for learning semantic-aligned features for enhanced action representation.
• To improve the robustness and discriminative capability of action recognition models against pose variations and occlusions.

Main Methods:

• Proposed a two-stage DCNN architecture: the first maps image regions to human body part response maps, and the second learns semantic-aligned features.
• Incorporated group (part) sparseness prior during feature learning, leveraging the knowledge that few parts drive an action.
• Employed an iterative mining method to learn discriminative action primitive detectors.

Main Results:

• The learned semantic-aligned features significantly boost action representation discriminability.
• The proposed method demonstrates robustness to pose variations and occlusions.
• Achieved superior recognition performance on action recognition benchmarks.

Conclusions:

• The developed DCNN framework effectively learns semantic-aligned features for robust and discriminative action recognition.
• This approach addresses the fundamental bottleneck of semantic misalignment in local action features.
• The findings suggest a promising direction for advancing human action understanding in computer vision.