Enhancing Dense Object Counting in Occlusion with a Dual-Branch Network

Object counting is a crucial technique that has wide-ranging applications in various domains. A significant challenge in this area is to accurately count dense objects in the presence of occlusions. Previous studies typically used single-branch networks to estimate target quantities, but they fell short in effectively managing occlusion issues within dense clusters. In this study, we introduced the Bilateral Counting Network (BCN), an innovative framework designed to overcome these limitations. BCN incorporates the Dense Region Extraction (DRE) algorithm, an image segmentation method based on clustering that efficiently partitions dense regions of objects by analyzing connected domains. Additionally, the Multi-Lateral Collaborative Counting Network (MCCN), a component of BCN, is a multibranch network that learns to process divided dense regions separately, reducing interference with the generalization features of less occluded areas. Our method performs well on dense small object counting tasks.