行人属性--HydraPlus-Net: Attentive Deep Features for Pedestrian Analysis

HydraPlus-Net: Attentive Deep Features for Pedestrian Analysis
ICCV2017
https://github.com/xh-liu/HydraPlus-Net

本文首次将 attention idea 应用到 行人属性分析上来。

行人分析的难度还是比较大，因为不同场合分析的侧重点有所不同，有时需要侧重局部信息，有时需要侧重全局信息。Semantic-level、 Low-level、 Scales

本文提出了一个网络结构 HydraPlus-Net Architecture

其中 Attentive Feature Net (AF-net) 包含三个 multidirectional attention (MDA) modules

multi-directional attention (MDA) 结构示意图如下所示：

接下来我们来看看什么是 attention idea？
从 CNN 网络不同的 blocks 学习到的 attention maps 会对不同的尺度和结构产生响应
It is well known that the attention maps learned from different blocks vary in scale and detailed structure

从高层 block 学习到的 attention maps 更多的侧重 semantic 区域，更粗糙些
the attention maps from higher blocks (e.g. α 3 ) tend to be coarser but usually figure out the semantic regions like α 3 highlights the hand bag in Fig.4(a)

从底层block 学习到的 attention maps 更多的侧重 细节信息如边缘纹理
those from lower blocks (e.g. α 1 ) often respond to local feature patterns and can catch detailed local information like edges and textures, just as the examples visualized in Fig. 4(a)

这里我们使用 MDA 模块将各种层次的信息融合起来，使其更有表征能力
if fusing the multi-level attentive features by MDA modules, we enable the output features to gather information across different levels of semantics, thus offering more selective representations

attention modules 越多效果越好

HP-net and DeepMar 对比

针对行人属性分析我们建立了一个新的数据库 PA-100K dataset

性能对比

Person Re-identification

11