Learning Match Kernels on Grassmann Manifolds for Action Recognition.

Area of Science:

• Computer Vision
• Machine Learning
• Pattern Recognition

Background:

• Action recognition is crucial for various applications but faces challenges in modeling and comparing actions.
• Existing methods struggle with capturing both local patterns and temporal dynamics of motion effectively.

Purpose of the Study:

• To propose a novel approach for action recognition using Grassmann match kernels (GMK).
• To develop a method for modeling actions as linear subspaces on the Grassmann manifold.
• To enhance the discriminative ability of action similarity measurement.

Main Methods:

• Actions are modeled as linear subspaces using temporally pooled convolutional neural network (CNN) features.
• Grassmann match kernels (GMK) are proposed, based on canonical correlations of linear subspaces.
• GMK is learned in a supervised manner via kernel target alignment for improved discrimination.

Main Results:

• The proposed GMK approach leverages CNNs for feature extraction and kernels for similarity measurement.
• Extensive experiments were conducted on five challenging datasets: Youtube, UCF50, UCF101, Penn Action, and HMDB51.
• The approach achieved high performance, significantly outperforming state-of-the-art algorithms.

Conclusions:

• The proposed GMK method offers a general learning framework for match kernels in action recognition.
• The approach demonstrates significant effectiveness and superiority over existing methods.
• This work advances the field of action recognition through improved subspace modeling and kernel learning.