GCNet: Probing self-similarity learning for Generalized Counting Network

Mingjie Wang, Yande Li, Jun Zhou, Graham W. Taylor, Minglun Gong

Abstract: Highlights•In use: A pseudo exemplar simulator is developed to automatically and effectively learn pseudo exemplar cues. From this, a self-similarity learning scheme is designed to capture a semantically-aware similarity map. It marries our method with the fully exemplar-free property during both training and inference phases while avoiding the burdensome two-stage training.•Data collection: Weakly-supervised regression eliminates the need for density maps obtained by labour-intensive point labels. This allows GCNet to be trained end-to-end using only a count-level supervisory signal. Together with the exemplar-free property, this makes GCNet require much less annotator effort.•Performance: The state-of-the-art accuracy is achieved on the prevailing benchmark FSC147 compared to the exemplar-free RepRPN-Counter and the BMNet-based baseline. Impressive results on FSC147 also demonstrate the superiority of our GCNet compared to traditional CAC approaches using both exemplars and location annotations.•Application scope: Repurposing our GCNet for challenging crowd-specific datasets (e.g., ShanghaiTech Part A, Part B and UCF_QNRF) further illustrates its strong generality.