Core Concepts
PSDiff formulates person search as a dual denoising process, achieving state-of-the-art performance through iterative and collaborative refinement.
Abstract
The content introduces PSDiff, a novel diffusion-based framework for person search. It addresses challenges in existing methods by eliminating prior pedestrian candidates and promoting collaboration between detection and ReID tasks. Extensive experiments on CUHK-SYSU and PRW datasets show superior performance with fewer parameters. The methodology includes a feature extractor, dual noise generator, and collaborative denoising layer. In-depth analysis of related works, diffusion model preliminaries, model architecture, training process, inference strategy, and comparisons with SOTA methods are provided.
Stats
PSDiffは、状態-of-the-artのパフォーマンスを達成します。
PSDiffは、CUHK-SYSUデータセットで95.1%のmAPを達成します。
PSDiffは、PRWデータセットで57.1%のmAPを達成します。