1607.Gated Siamese Convolutional Neural Network Architecture for Human Re-identificationl论文笔记

作者的贡献:
1607.Gated Siamese Convolutional Neural Network Architecture for Human Re-identificationl论文笔记
作者针对Siamese network结构进行改进,采用更小的卷积核和更深的架构,包括采用最后几层采用更小的不对称的卷积核,引入BN层加快网络收敛,更重要的是,作者在最后四层之间插入matching gate层,通过比较每一特征图的每一排,在水平条纹上生成基于同一ID的更底层的特征掩膜,并将同一水平区域的有效区域特征整合到最终每层的特征提取中,作者将这种新颖的架构最终命名为Gated S-CNN网络,如下所示:
1607.Gated Siamese Convolutional Neural Network Architecture for Human Re-identificationl论文笔记
1607.Gated Siamese Convolutional Neural Network Architecture for Human Re-identificationl论文笔记

作者在market-1501,cuhk03,VIPeR数据集上做了测试,如下:
1607.Gated Siamese Convolutional Neural Network Architecture for Human Re-identificationl论文笔记
1607.Gated Siamese Convolutional Neural Network Architecture for Human Re-identificationl论文笔记
1607.Gated Siamese Convolutional Neural Network Architecture for Human Re-identificationl论文笔记
另外,作者发布的补充材料中,含有的内容有:作者采用控制变量法验证所加各结构的有效性,一些超参数的设置说明及含有gate的孪生网络梯度反传公式推导。示例如下:
1607.Gated Siamese Convolutional Neural Network Architecture for Human Re-identificationl论文笔记

1607.Gated Siamese Convolutional Neural Network Architecture for Human Re-identificationl论文笔记

1607.Gated Siamese Convolutional Neural Network Architecture for Human Re-identificationl论文笔记