Title: One-Shot Crowd Counting With Density Guidance For Scene Adaptation

Abstract: Surveillance footage captured from various vantage points exhibits significant variability, yet current crowd counting models often struggle to generalize effectively to unseen monitoring environments. To enhance model adaptability, we treat distinct surveillance settings as separate categories and employ few-shot learning techniques to enable the system to adjust to new scenes based on provided exemplar categories. Central to our approach is the utilization of both local and global density traits to steer the counting process for these unfamiliar environments. Specifically, to accommodate the diverse density fluctuations found in target scenes, we introduce a multiple local density learner. This component identifies multi-prototypes that capture various density distributions present in the support scene. These local density similarity matrices are subsequently encoded and applied to guide the model through localized adjustments. Additionally, to align with the overall density patterns of the target scene, we extract global density features from the support image, which are then used to direct the model on a global scale. Evaluations across three surveillance datasets demonstrate that our proposed method successfully adapts to unseen monitoring scenarios, surpassing recent state-of-the-art techniques in the realm of few-shot crowd counting.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC