What-and-where to match: Deep spatially multiplicative integration networks for person re-identification

作者：

Highlights：

• A novel deep architecture to emphasize common local patterns is proposed to learn flexible joint representations for person re-identification.

• The proposed method introduces a multiplicative integration gating function to embed two convolutional features to their joint representations, which are effective in discriminating positive pairs from negative pairs.

• Spatial dependencies are incorporated into feature learning to address the cross-view misalignment.

• Extensive experiments and empirical analysis are provided in experimental part.

摘要

•A novel deep architecture to emphasize common local patterns is proposed to learn flexible joint representations for person re-identification.•The proposed method introduces a multiplicative integration gating function to embed two convolutional features to their joint representations, which are effective in discriminating positive pairs from negative pairs.•Spatial dependencies are incorporated into feature learning to address the cross-view misalignment.•Extensive experiments and empirical analysis are provided in experimental part.

论文关键词：Multiplicative integration gating,Convolutional neural networks,Recurrent neural networks,Person re-identification

论文评审过程：Received 17 May 2017, Revised 20 September 2017, Accepted 6 October 2017, Available online 13 October 2017, Version of Record 8 January 2018.

论文官网地址：https://doi.org/10.1016/j.patcog.2017.10.004