A2SNet: An Adaptive Scale Selection Network for Crowd Counting in the Wild

As deep learning is widely used in diverse computer vision applications, the study of crowd counting has achieved superhuman performance. However, current methods tend to directly concatenate multi-scale features from different branches to reduce the error of counting. Such an approach hurt the multi-scale feature representation learning, therefore hard to distinguish the priority of different features. Different from previous studies, we propose a novel Adaptive Scale Selection Network (A2SNet) that learns multi-scale features to adaptively balance weights. Our approach contains multiple branches: one branch uses Adaptive Scale Selection (ASS) module for multi-scale features extraction, and the second branch employs Efficient Channel Attention (ECA) module to further learn subtle features. The proposed ASS is a parallel structure, which consists of coarse branch (i.e., large-scale learning) and fine branch (i.e., small-scale learning). Lastly, we reveal A2SNet’s effectiveness on standard crowd benchmarks and show remarkable results.