MVReID-Former: Multi-View Vision Transformer for Cross-Camera Person and Vehicle Re-Identification in Surveillance Networks

Cross-camera person and vehicle re-identification in surveillance networks remains a challenging task due to variations in camera viewpoints, image quality, and environmental conditions. Traditional methods using hand-crafted features show limited performance in real-world scenarios. We propose MVReID-Former, a novel Multi-View Vision Transformer architecture specifically designed for cross-camera re-identification in surveillance networks. Our approach introduces three key innovations: (1) a Cross-View Attention Mechanism that learns discriminative features across different camera viewpoints and qualities, (2) Camera-Aware Positional Encoding that incorporates camera metadata including resolution, vantage point, and quality into attention computation, and (3) Adaptive Quality Weighting that handles varying image quality from VGA to 3MP resolutions. Experiments on the Dana36 dataset demonstrate improvements over existing methods, with our approach showing promising results for multi-camera surveillance scenarios. The hierarchical transformer design effectively captures both finegrained patch-level details and global cross-view correspondences, enabling robust identity matching across challenging scenarios.