Revisiting Multi-scale Neural Network for Crowd Counting
In crowd counting tasks,there are still issues such as perspective distortions and crowd distribution variations.Multi-scale architecture in deep neural networks(DNNs)is employed to solve these problems.Feature extracted by multi-scale architecture can be either directly fused or fused through the guidance of proxies in DNNs.However,these fusion methods are not capable of dealing with the per-pixel performance discrepancy over multi-scale density maps.Therefore,an expert system is introduced to hierarchically fuse the multi-scale density maps through the pixel-level soft weights obtained from the pixel-level gating network.A competition-cooperation strategy is also proposed to ensure that all experts from all scales can work.Experiments on some public datasets show that the proposed method achieves the state-of-the art performance.
crowd countingmulti-scale neural networkmixture of experts
NETL
NSTL
万方数据
