Focal Inverse Distance Transform Maps for Crowd Localization

In this paper, we focus on the crowd localization task, a crucial topic of analysis. Most regression-based methods utilize convolution neural networks (CNN) to regress density map, which can not accurately locate instance in extremely dense scene, attributed two reasons: 1) map consists series blurry Gaussian blobs, 2) severe overlaps exist region map. To tackle issue, propose novel Focal Inverse Distance Transform (FIDT) for task. Compared with maps, FIDT maps describe persons' locations without overlapping regions. Based Local-Maxima-Detection-Strategy (LMDS) is derived effectively extract center point each individual. Furthermore, introduce an Independent SSIM (I-SSIM) loss make model tend learn local structural information, better recognizing maxima. Extensive experiments demonstrate that proposed method reports state-of-the-art performance six datasets and one vehicle dataset. Additionally, find shows superior robustness negative scenes, further verifies effectiveness maps. The code will be available at https://github.com/dk-liang/FIDTM.