Pose-Guided person re-identification using attention-enhanced GANs with feature fusion
Abstract
Pose variation remains a major challenge in person re-identification (ReID), particularly in unconstrained surveillance environments with non-overlapping camera views. This paper presents a novel pose-guided ReID framework that integrates an attention-enhanced generative adversarial network (GAN) with a feature fusion strategy to address this challenge. The proposed architecture synthesizes pose-aligned pedestrian images using a conditional autoencoding GAN, which preserves identity while normalizing viewpoint discrepancies. Identity features are then extracted using a w-50 backbone augmented with spatial and channel attention modules. A lightweight fusion module adaptively combines features from both real and generated images, enhancing the robustness and discriminability of the final embedding. Extensive evaluations on Market-1501 and DukeMTMC-reID datasets demonstrate the effectiveness of our approach, achieving 89.6% Rank-1 accuracy and 74.3% mAP, outperforming several state-of-the-art models. These results confirm that the integration of pose-guided synthesis, attention-driven feature extraction, and learned fusion significantly improves ReID performance under cross-view and occlusion conditions.