Uniform and Variational Deep Learning for RGB-D Object Recognition and Person Re-Identification

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Traditional object recognition and person re-identification heavily rely on RGB images, limiting robustness to viewpoint variations.
• Exploiting geometric and anthropometric information from RGB-D data offers a more reliable approach to recognition tasks.

Purpose of the Study:

• To propose a novel Uniform and Variational Deep Learning (UVDL) method for RGB-D object recognition and person re-identification.
• To enhance recognition accuracy by integrating geometric and anthropometric information with visual appearance.
• To develop a method robust to different viewpoints by utilizing RGB-D data.

Main Methods:

• Extracting depth and appearance features using two separate deep convolutional neural networks.
• Designing a uniform and variational multi-modal auto-encoder to project features into a common latent space.
• Jointly optimizing the auto-encoder and convolutional neural networks with discriminative loss and reconstruction error.

Main Results:

• Demonstrated the efficiency of the UVDL method on RGB-D object recognition tasks.
• Validated the effectiveness of the proposed approach for RGB-D person re-identification.
• Achieved improved recognition performance by effectively combining depth and appearance features.

Conclusions:

• The proposed UVDL method effectively utilizes RGB-D data for robust object recognition and person re-identification.
• The multi-modal auto-encoder successfully creates a discriminative latent space from combined features.
• This approach offers a significant advancement in leveraging multi-modal sensor data for recognition tasks.