A multi-scale network with density-guided structural similarity loss and adaptive local maxima detection for crowd counting

Crowd counting focuses on providing the number of people in an image, and it holds great application value in crowd management and public safety. Most existing methods generate the predicted map during the training phase and obtain the number of people through post-processing in the inference phase. However, these methods do not consider the multi-scale heads, i.e., a large variation of head size in the image, both in the loss function of the training phase and the post-processing of the inference phase. In this paper, we propose a novel Multi-scale Network with Density-guided Structural Similarity (DSSIM) Loss and Adaptive Local Maxima Detection (ALMD) for Crowd Counting to alleviate the multi-scale heads. During the training phase, DSSIM loss is calculated within the head region. To this end, an adaptive window is designed for each head based on crowd density, enabling precise delineation of the head region. During the inference phase, we estimate the number of people by setting the adaptive thresholds in the proposed post-processing mechanism ALMD. Specifically, we classify the image patches into three levels based on crowd density and assign different adaptive thresholds to the three levels to accurately filter out head points. Extensive experiments on three public datasets confirm the effectiveness of our approach.