Distance Metric Learning Using Privileged Information for Face Verification and Person Re-Identification

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Face verification and person re-identification commonly rely solely on RGB images.
• Depth information from RGB-D cameras offers potential for improved performance but is often underutilized.

Purpose of the Study:

• To develop a novel approach for enhancing face verification and person re-identification using RGB-D data.
• To leverage depth features as privileged information during the training of distance metric learning models.

Main Methods:

• Proposed a distance metric learning framework incorporating privileged depth information (ITML+).
• Utilized depth features from RGB-D data (e.g., Kinect) during training.
• Developed an efficient algorithm based on cyclic projection for the ITML+ formulation.

Main Results:

• Demonstrated significant improvements in face verification and person re-identification accuracy.
• Validated the effectiveness of using depth features as privileged information.
• Achieved state-of-the-art or competitive results on benchmark datasets (EUROCOM, CurtinFaces, BIWI RGBD-ID).

Conclusions:

• Integrating depth features as privileged information in distance metric learning effectively enhances RGB-based face verification and person re-identification.
• The proposed ITML+ method and its efficient algorithm offer a robust solution for leveraging RGB-D data.