Dynamic proxy domain generalizes the crowd localization by better binary segmentation

Crowd localization aims to predict the precise location of each instance within an image. Current advanced methods utilize pixel-wise binary classification to address the congested prediction, where pixel-level thresholds convert prediction confidence into binary values for identifying pedestrian heads. Due to the extremely variable contents, counts, and scales in crowd scenes, the confidence-threshold learner is fragile and lacks generalization when encountering domain shifts. Moreover, in most cases, the target domain is unknown during training. Therefore, it is crucial to explore how to enhance the generalization of the confidence-threshold locator to latent target domains. In this paper, we propose a Dynamic Proxy Domain (DPD) method to improve the generalization of the learner under domain shifts. Concretely, informed by the theoretical analysis of the upper bound of generalization error risk for a binary classifier on latent target domains, we introduce a generated proxy domain to facilitate generalization. Then, based on this theory, we design a DPD algorithm consisting of a training paradigm and a proxy domain generator to enhance the domain generalization of the confidence-threshold learner. Additionally, we apply our method to five types of domain shift scenarios, demonstrating its effectiveness in generalizing crowd localization. Our code is available at https://github.com/zhangda1018/DPD.