Cross-View Discriminative Feature Learning for Person Re-Identification.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Viewpoint variability across non-overlapping cameras poses a significant challenge for person re-identification (Re-ID) systems.
• Existing methods struggle to effectively handle cross-view ambiguity, limiting Re-ID performance.

Purpose of the Study:

• To mitigate cross-view ambiguity in person Re-ID by learning highly discriminative deep features.
• To develop a novel loss function that improves the learning of intra-class and inter-class relationships in the feature domain.

Main Methods:

• A novel loss function comprising a steering meta center term and an enhancing centers dispersion term was proposed.
• This objective function guides the training process to mine effective relationships within and between identity features.
• The approach jointly learns embeddings and the metric contextually, optimizing feature space expansion and class compactness.

Main Results:

• The proposed loss supervision generates an expanded feature space with compact classes, reducing inter-identity interference.
• The technique enhances Convolutional Neural Network (CNN) invariance to viewpoint without increased training complexity.
• State-of-the-art performance was achieved on the Market-1501 and CUHK03 benchmark datasets.

Conclusions:

• The novel loss function effectively addresses viewpoint variability in person Re-ID.
• The method offers a more optimized approach to learning discriminative features compared to existing metric learning techniques.
• This work advances the field of person Re-ID by improving viewpoint invariance and overall performance.