Towards More Powerful Multi-column Convolutional Network for Crowd Counting

Jiabin Zhang, Qi Chu, Weihai Li, Bin Liu, Weiming Zhang, Nenghai Yu

Abstract: Scale variation has always been one of the most challenging problems for crowd counting. By using multi-column convolutions with different receptive fields to deal with different scales in the scene, the multi-column convolutional networks have achieved good performance. However, there is still great potential waiting to be explored for multi-column convolutional networks. To this end, we propose to design a multi-column neural network that can more effectively adapt to scene scale variations automatically, by applying Neural Architecture Search technology. First, we combine Progressive Neural Architecture Search scheme with crowd counting to construct our Progressive Multi-column Architecture Serach (PMAS) framework. Furthermore, to reduce the bias caused by the weight-share scheme, which is widely adopted in efficient Neural Architecture Search, we propose a novel pre-architecture-based weight-share scheme. Experiments on several challenging datasets demonstrate the effectiveness of our method.