Improving crowd counting with scale‐aware convolutional neural network

Abstract

Large-scale variations may cause a serious problem in crowd counting. In recent years, most methods for this problem use convolutional neural networks with a fixed scale for encoding and decoding image features. The scale of the convolutional layer is usually manually adjusted and may have to deal with image features on unfitted scales. In this paper, a method called scale-aware convolutional neural network(SCNet) is proposed, which adds a scale selection mechanism to the dilated convolutional operation. Shared weight multi-branch is used to deal with features on different scales, and an attention mechanism is introduced to determine the weights of the branches that fit the scale. Experimental results demonstrate that the proposed SCNet outperforms most existing methods.