Discriminative Multi-View Dynamic Image Fusion for Cross-View 3-D Action Recognition

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Depth video action recognition faces challenges due to viewpoint variation.
• Existing methods struggle with view-tolerance while maintaining feature distinctiveness.

Purpose of the Study:

• To enhance view-tolerance and discriminative capacity for 3-D action recognition in depth videos.
• To propose a discriminative fusion method for Multi-view Dynamic Images (MVDI) using Multi-Instance Learning (MIL).

Main Methods:

• Representing 3-D actions using Multi-view Dynamic Images (MVDI).
• Encoding MVDI instances with Fisher Vectors (FV) for higher dimensional mapping.
• Aggregating FV-encoded instances via sum-pooling for a robust action signature.
• Implementing a discriminative viewpoint instance discovery to discard irrelevant views.

Main Results:

• The proposed method significantly improves cross-view 3-D action recognition performance.
• The approach demonstrates effectiveness across five diverse datasets.
• The method is also applicable to cross-view 3-D object recognition tasks.

Conclusions:

• The discriminative MVDI fusion method effectively enhances view-tolerance in 3-D action recognition.
• Fisher Vectors and MIL provide a powerful framework for robust action representation.
• The proposed technique offers a significant advancement for recognizing actions from depth videos with varying viewpoints.