Crowd estimation in images

Summarise with VIRAT
Authors: Matthias Xing Ran Lim
Cited by: 0

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Crowd counting is an essential component for public safety and urban management, as its applications range from disaster prevention to monitoring social distancing. Accordingly, this project investigates the effectiveness of including a feature-level spatial gating mechanism into the Point Query Transformer (PET) framework to improve crowd counting performance. For this purpose, the proposed modification introduces a lightweight spatial gating module between the CNN backbone and the progressive encoder attention module. This is designed so that it suppresses background noise while also emphasizing crowd-relevant spatial features. The modified architecture was then evaluated using three benchmark datasets, namely ShanghaiTech Part A and B as well as UCF-QNRF, with Mean Absolute Error (MAE) and Root Mean Squared Error (RMSE) serving as evaluation metrics. The results from running the model based on the above architecture indicate that the spatial gating mechanism improves performance in lower-density areas, achieving a reduced MAE on ShanghaiTech Part B. Conversely, performance degradation is observed in dense crowd scenarios due to over-suppression of crowd-related features. Qualitative analysis further reveals limitations of the model, including false attention on human-like background objects as well as a mismatch between gate activation and prediction output. As such, these findings highlight both the potential and current limitations of spatial gating as a strategy for refining crowd counting feature refinement. It likewise provides a foundation for future work on adaptive and semantically-aware gating mechanisms.

Fields of Study & Subjects

Video Surveillance and Tracking Methods Advanced Neural Network Applications Visual Attention and Saliency Detection

Publication details

Journal / source
DR-NTU (Nanyang Technological University)
Publication date
1 January 2026
Language
EN