Dynamic frequency selection and spatial interaction fusion for robust person search

Abstract

Person search aims to locate target individuals in large image databases captured by multiple non-overlapping cameras. Existing models primarily rely on spatial feature extraction to capture fine-grained local details, which is vulnerable to background clutter and occlusions and leads to unstable feature representations. To address the issues, we propose a Dynamic Frequency Selection and Spatial Interaction Fusion Network (PS-DFSI), marking the first attempt to introduce frequency decoupling and selection into person search. By integrating frequency and spatial features, PS-DFSI enhances feature expressiveness and robustness. Specifically, it comprises two core modules: the Dynamic Frequency Selection Module (DFSM) and the Spatial Frequency Interaction Module (SFIM). DFSM decouples feature maps into low-frequency and high-frequency components using learnable low-pass and high-pass filters, and a frequency selection modulator emphasizes key frequency components via channel attention. SFIM refines local details by fusing frequency-enhanced features with high-level semantic representations, leveraging multi-scale receptive fields and cross-feature attention for efficient spatial-frequency integration. Extensive experiments on CUHK-SYSU and PRW demonstrate that PS-DFSI significantly improves person search performance, validating its effectiveness and robustness.