PS-ARM: An End-to-End Attention-Aware Relation Mixer Network for Person Search

Person search is a challenging problem with various real-world applications, that aims at joint person detection and re-identification of query from uncropped gallery images. Although, previous study focuses on rich feature information learning, it’s still hard to retrieve the due occurrence appearance deformations background distractors. In this paper, we propose novel attention-aware relation mixer (ARM) module for search, which exploits global between different local regions within RoI make it robust against occlusion. The proposed ARM composed block spatio-channel attention layer. introduces spatially attended spatial mixing channel-wise channel effectively capturing discriminative features an RoI. These are further enriched by introducing where foreground discriminability empowered in space. Our generic does not rely fine-grained supervisions or topological assumptions, hence being easily integrated into any Faster R-CNN based methods. Comprehensive experiments performed two benchmark datasets: CUHK-SYSU PRW. PS-ARM achieves state-of-the-art performance both datasets. On PRW dataset, our absolute gain 5% mAP score over SeqNet, while operating comparable speed. source code pre-trained models available https://github.com/mustansarfiaz/PS-ARM .