Hierarchical Clustering Multi-Task Learning for Joint Human Action Grouping and Recognition.

Area of Science:

• Computer Vision
• Machine Learning
• Pattern Recognition

Background:

• Human action recognition and grouping are crucial in various applications.
• Existing methods often rely on heuristic or subjective action grouping, leading to inconsistencies.
• Previous multi-task learning approaches assume actions are either independent or fully correlated.

Purpose of the Study:

• To propose a novel Hierarchical Clustering Multi-Task Learning (HC-MTL) method for joint human action grouping and recognition.
• To develop an automated approach that overcomes the limitations of heuristic action grouping.
• To enable joint optimization of action models and latent group discovery.

Main Methods:

• Formulated an objective function using group-wise least square loss, regularized by low rank and sparsity.
• Decomposed the non-convex optimization into multi-task learning and task relatedness discovery sub-tasks.
• Iteratively alternated between convex formulation for multi-task learning and leveraging learned parameters for clustering.

Main Results:

• HC-MTL achieved competitive performance against state-of-the-art methods in action recognition and grouping.
• Demonstrated the ability to overcome heuristic action grouping difficulties and inconsistencies.
• Showcased improved performance compared to clustered multi-task learning by effectively utilizing discovered latent relatedness.

Conclusions:

• HC-MTL provides an effective and automated solution for joint human action grouping and modeling.
• The method successfully breaks the assumption of independent or fully correlated actions in multi-task learning.
• HC-MTL offers a more objective and robust approach to action grouping based on feature subspace distributions.