earticle

영어

In this paper, we propose a new framework that enables an object detector trained with only point-level annotations to estimate the centroids and sizes of objects in dense scenes. Specifically, the framework is based on the Swin Transformer structure and introduces a self-designed resolution feature fusion module in the hierarchical structure, where the estimation of object centroids is done directly by point supervision, and the object pseudo-size is initialized based on the assumption of local uniform distribution, and the regression of object size is guided by an improved congestion-aware loss function. In the NWPU-Crowd dataset, our method outperformed the existing state-of-the-art detection counting methods in F1-measure, precision, MSE evaluation criteria.