Crowd Counting with Segmentation Map Guidance
Hao Xu, Chengyao Zheng, Yuncong Nie, Siyu Xia · 2019
As one of the fundamental vision tasks, crowd counting has attracted a tremendous amount of efforts and achieved significant improvement over the decades. To be applied in real-world applications, the techniques of crowd counting need to be robust and efficient. However, the state-of-the-art works may easily fail in front of an extremely complex background. The reason is that the background is very similar to the small congested heads under some circumstances. To this end, we propose a new deep model to integrate a segmentation map to compensate for the false response under a complex environment. Additionally, we propose a new method to generate segmentation ground truth merely based on the density map instead of manual labeling. Besides, we modify the geometry-adaptive method for density generation employed in the highly congested area. Experiments on ShanghaiTech dataset demonstrate the advantage of the proposed models over the state-of-the-art.