Multi-Domain & Multi-Task Learning for Human Action Recognition.

Area of Science:

• Computer Vision
• Machine Learning
• Human-Computer Interaction

Background:

• Domain-invariant feature representation (view-invariant and modality-invariant) is crucial for robust human action recognition.
• Discovering latent correlations among multiple actions is essential for effective action modeling.

Purpose of the Study:

• To propose a multi-domain and multi-task learning (MDMTL) method for human action recognition.
• To extract domain-invariant information for multi-view and multi-modal action representation.
• To explore the relatedness among multiple action categories for improved action modeling.

Main Methods:

• A sparse transfer learning-based method is used to co-embed multi-domain data into a common space for discriminative feature learning.
• Visual feature learning is integrated within a multitask learning framework.
• Frobenius-norm regularization and sparse constraints are employed for joint task modeling and task relatedness-induced feature learning.

Main Results:

• The proposed MDMTL method achieves superior performance compared to state-of-the-art approaches.
• Experiments demonstrate the effectiveness of the framework on benchmark datasets like IXMAS, DailyActivity3D, and M2I.
• The method successfully extracts domain-invariant features and models action relatedness.

Conclusions:

• MDMTL is the first supervised framework to jointly achieve domain-invariant feature learning and task modeling for multi-domain action recognition.
• The approach offers a significant advancement in multi-view and multi-modal human action recognition.
• The method's effectiveness is validated across diverse and challenging datasets.