Crowd Counting via Segmentation Guided Attention Networks and Curriculum Loss

Automatic crowd behaviour analysis is an important task for intelligent transportation systems to enable effective flow control and dynamic route planning varying road participants. Crowd counting one of the keys automatic analysis. using deep convolutional neural networks (CNN) has achieved encouraging progress in recent years. Researchers have devoted much effort design variant CNN architectures most them are based on pre-trained VGG16 model. Due insufficient expressive capacity, backbone network usually followed by another cumbersome specially designed good performance. Although VGG models been outperformed Inception image classification tasks, existing built with modules still only a small number layers basic types modules. To fill this gap, paper, we firstly benchmark baseline Inception-v3 model commonly used datasets achieve surprisingly performance comparable or better than models. Subsequently, push boundary disruptive work further proposing Segmentation Guided Attention Network (SGANet) as novel curriculum loss counting. We conduct thorough experiments compare our SGANet prior arts proposed can state-of-the-art MAE 57.6, 6.3 87.6 ShanghaiTechA, ShanghaiTechB UCF_QNRF, respectively.