Deep Manifold Learning Combined With Convolutional Neural Networks for Action Recognition.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Deep learning is widely used for action recognition.
• Utilizing structural manifold information among action videos can improve recognition accuracy and efficiency.
• Few studies have explored manifold information for action recognition.

Purpose of the Study:

• To propose a novel Deep Manifold Learning (DML) framework.
• To enhance action recognition accuracy and efficiency by incorporating training sample manifold information.
• To adapt DML to existing deep learning architectures.

Main Methods:

• The proposed DML framework integrates manifold information into deep learning models.
• For convolutional neural networks, DML embeds manifold information between convolutional layers to boost feature discriminability.
• DML is also applied to Restricted Boltzmann Machines to mitigate overfitting.

Main Results:

• Experimental results were validated on four standard action recognition datasets: UCF101, HMDB51, KTH, and UCF sports.
• The DML framework demonstrated superior performance compared to existing state-of-the-art methods.
• DML effectively enhances the discriminative capacity of deep learning layers.

Conclusions:

• Deep Manifold Learning (DML) is a viable approach to improve action recognition.
• The DML framework offers a flexible way to enhance various deep learning models.
• This method significantly advances the state-of-the-art in action recognition research.