Unsupervised Hierarchical Dynamic Parsing and Encoding for Action Recognition.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Action evolution in videos is often non-uniform, featuring complex rhythms and non-stationary dynamics.
• Existing methods struggle to model these intricate temporal variations effectively.

Purpose of the Study:

• To develop a novel hierarchical dynamic parsing and encoding method for action recognition.
• To capture both locally smooth dynamics and globally drastic dynamic changes within action sequences.

Main Methods:

• A two-layer hierarchical approach is proposed: unsupervised parsing into smooth-changing stages using temporal clustering, followed by encoding within stages using mean-pooling or rank-pooling.
• The second layer encodes ordered dynamics from the first layer using rank-pooling for a joint representation.

Main Results:

• The method effectively models non-uniform temporal dynamics in action sequences.
• Demonstrated superior performance on gesture (Chalearn Gesture) and generic action datasets (Olympic Sports, Hollywood2, UCF101).

Conclusions:

• The proposed hierarchical dynamic parsing and encoding method enhances action recognition accuracy.
• This approach provides a robust framework for analyzing complex temporal patterns in video data.