EESSO: Exploiting Extreme and Smooth Signals via Omni-frequency learning for Text-based Person Retrieval

Jingyi Xue, Zijie Wang, Guannan Dong, Aichun Zhu

Abstract: Highlights•A jointly optimized multi-stream architecture named EESSO is proposed.•A Frequency Component Decomposition Module is utilized to separate high-low signal.•An Uncertainty-Guided Mutual Learning Mechanism is designed.•Address the task of text-based person retrieval from the frequency perspective.•Experimental analysis demonstrates that EESSO achieves the SOTA performance.