Crowd Counting Using Scale Enhanced Network with Dual Attention Booster

Crowd counting has been a fundamental yet challenging problem in pattern recognition. Most recent deep models for crowd counting rely on Convolutional Neural Networks (CNNs). Although CNN visual features comprise the spatial and channel features, existing deep models on crowd counting have limited descriptive ability as they only focus on the spatial or channel information. In this paper, we propose Scale Enhanced Network with Dual Attention Booster named as SEN-DAB, a novel method to jointly learn the representations of spatial and channel information for crowd counting. Moreover, to further leverage the multi-scale information, a pyramid residual scale enhanced block is presented to process the multi-scale features. As a result, the learned spatial, channel and multi-scale features can be robust to appearance changes of the crowd. Our model is tested on three benchmarks and the experimental results confirm that the promising performance of SEN-DAB when compared with various networks.