[ECCV2018]Person Search via A Mask-Guided Two-Stream CNN Model

PDF

这篇文章和以往的person search文章不同,作者提出detector和re-ID分离才会产生更好的performance。detector是faster RCNN产生proposal和离线(off-the-shelf)的instance segmentation method FCIS产生分割的mask;re-ID是用的Two-stream CNN完成。

Contribution

[ECCV2018]Person Search via A Mask-Guided Two-Stream CNN Model

Method

[ECCV2018]Person Search via A Mask-Guided Two-Stream CNN Model

框架如上图。Detector是用的faster RCNN,分割是用的现有的算法FCIS;随后对原图像块和前景图像分别输入O-Net和F-Net,然后用concatenate将其整合在一起。再利用SEBlock对权重进行再分配,在经过GAP(Global Average Pooling)和FC,对输出向量求OIM loss。

SEBlock

[ECCV2018]Person Search via A Mask-Guided Two-Stream CNN Model

OIM Loss

[ECCV2018]Person Search via A Mask-Guided Two-Stream CNN Model

实验结果

[ECCV2018]Person Search via A Mask-Guided Two-Stream CNN Model

[ECCV2018]Person Search via A Mask-Guided Two-Stream CNN Model

 

上一篇:vector 定义的二维数组的遍历


下一篇:[读书心得]Runtime Type Fundamentals from CLR via C#