Towards Robustness Against Common Corruptions for Text-based Person Search

Text-based person search (TBPS) retrieves images of individuals based on natural language descriptions. This technology is frequently deployed in complex environments such as video surveillance, where it is susceptible to various visual corruptions, including snow, frame compression, and motion blur. Current TBPS models lack robustness studies against these corruptions. This paper examines the impact of different types of corruption on TBPS models through frequency domain analysis. We introduce two innovative approaches to enhance robustness: hybrid adversarial training and image distribution adaptation. Hybrid adversarial training merges conventional and noise-based adversarial training methods, bolstering the model's resistance to high-frequency corruptions during the training phase. Concurrently, image distribution adaptation adjusts to the low-frequency characteristics of corrupted images, reconciling distribution discrepancies between training and corrupted testing datasets. By integrating these strategies, our enhanced models achieve significant improvements, with at least a 2.02% increase in average Rank@1 accuracy in common corruption tests on the CUHK-PEDES dataset.