2021, IEEE TIP, Point-Supervised Crowd Detection and Counting
In this article, we propose a novel self-training approach named Crowd-SDNet that enables a typical object detector trained only with point-level annotations (i.e., objects are labeled with points) to estimate both the center points and sizes of crowded objects. This is achieved by the proposed locally-uniform distribution assumption, the crowdedness-aware loss, the confidence and order-aware refinement scheme, and the effective decoding method, which promote the detector to generate accurate bounding boxes in a coarse-to-fine and end-to-end manner.
Article
Yi Wang, Junhui Hou, Xinyu Hou, and Lap-Pui Chau*. “A Self-Training Approach for Point-Supervised Object Detection and Counting in Crowds.” IEEE TIP 2021.
Article Preprint Supplemental Information GitHub Repo Poster