Grafted network for person re-identification

作者：

Highlights：

• GraftedNet grafts a high-accuracy network and a light-weighted network.

• Joint multi-level and part-based feature can improve person feature representation.

• Accompanying branch helps GraftedNet to train better; can be removed in testing.

• GraftedNet has better performance than original ResNet-50 with 4.6M parameters.

摘要

•GraftedNet grafts a high-accuracy network and a light-weighted network.•Joint multi-level and part-based feature can improve person feature representation.•Accompanying branch helps GraftedNet to train better; can be removed in testing.•GraftedNet has better performance than original ResNet-50 with 4.6M parameters.

论文关键词：Person re-identification,Feature representation,Multi-level feature,Part-based feature,Grafting

论文评审过程：Received 7 May 2019, Revised 21 October 2019, Accepted 23 October 2019, Available online 25 October 2019, Version of Record 31 October 2019.

论文官网地址：https://doi.org/10.1016/j.image.2019.115674