Reliable Pseudo-Supervision for Unsupervised Domain Adaptive Person Search.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Unsupervised Domain Adaptation (UDA) person search adapts models to new, unlabeled datasets.
• Current methods struggle with unreliable pseudo-supervision due to spectral shifts and static proxy updates.
• These limitations hinder feature stability and adaptability in person search models.

Purpose of the Study:

• To develop a novel framework, Reliable Pseudo-supervision in UDA Person Search (RPPS), to enhance pseudo-supervision reliability.
• To address spectral shift bias and static proxy updates in UDA person search.
• To improve the robustness and performance of person search models in cross-domain scenarios.

Main Methods:

• Proposed the Reliable Pseudo-supervision in UDA Person Search (RPPS) framework.
• Introduced a Dual-branch Wavelet Enhancement Module (DWEM) using Discrete Wavelet Transform (DWT) for feature decomposition and enhancement.
• Developed a Dynamic Confidence-weighted Clustering Proxy (DCCP) with confidence-guided initialization and a two-stage update strategy.

Main Results:

• RPPS achieved state-of-the-art performance on the CUHK-SYSU and PRW benchmarks.
• Demonstrated strong robustness against domain shifts.
• The proposed DWEM and DCCP modules significantly improved cross-domain robustness and discriminability.

Conclusions:

• Enhancing pseudo-supervision reliability is crucial for effective UDA person search.
• The RPPS framework offers a robust solution for cross-domain person search challenges.
• The findings highlight the potential of wavelet transforms and dynamic proxy updates in domain adaptation.