自适应感受野网络的行人重识别

行人重识别通常删除特征提取网络中的最后一个空间下采样操作,以增加最后输出特征图的分辨率,保留更多的细粒度特征.然而,这种操作会大幅减小神经网络的感受野,而更大的感受野可以为行人重识别提供更多的上下文信息.同时,在实际的视觉皮层中,相同区域的神经元的感受野是不同的,但当前行人重识别网络的设计大多忽视了这一点.为了解决上述问题,提出一种新颖的自适应感受野网络.网络的设计受启发于生物的视觉系统,通过在多分支网络上设置不同大小的感受野,结合注意力机制让网络自行选择合适的感受野特征,从而实现网络感受野的自适应,并且采用分组卷积使得自适应感受野模块更加轻量级.同时在各个分支利用空洞卷积增大感受野,补偿删除最后下采样操作所减少的网络感受野.在公开的大规模数据集上进行实验,实验结果表明,所提出的算法相比于基线方法有显著的提升,当使用ResNet-50作为特征提取网络时,在DukeMTMC-reID、Market-1501数据集上的Rank-1和mAP分别达到89.2%和76.0%、95.2%和87.2%.与现有方法相比,所提出算法在精度上有明显的提升.

Adaptive receptive network for person re-identification

Person re-identification typically removes the last spatial down-sampling operation in the backbone to increase the resolution of the final output feature map and preserve more fine-grained features. However, this operation substantially reduces the size of a receptive field, and a larger receptive field can provide more contextual information for person re-identification. At the same time, in the actual visual cortex, the receptive field of neurons in the same region are different, but this is largely ignored by the current design of pedestrian recognition networks. To solve the above problems, this article proposes a novel adaptive receptive field network. The design of the network is inspired by the visual system of living organisms. By setting a different sized receptive field on the multi-branch network, combined with the attention mechanism to allow the network to select the appropriate receptive field characteristics, the network receptive fields adaptive is realized, and the use of packet convolution makes the adaptive receptive field module more lightweight. The receptive field is also increased in each branch using empty convolution to compensate for the reduction of the network receptive field by deleting the last downsampling operation. Experiments are performed on publicly available large-scale datasets, and results show that the algorithm has a significant improvement over the baseline approach, with Rank-1 and mAP on the DukeMTMC-reID, Market-1501 datasets reaching 89.2% and 76.0%, 95.2% and 87.2%, respectively, when using ResNet-50 as backbone. Compared with the existing methods, the proposed algorithm also has a significant improvement in accuracy.