论文阅读: Omni-Scale Feature Learning for Person Re-Identification

pdf:
code:https://github.com/KaiyangZhou/deep-person-reid

Abstract:

As an instance-level recognition problem, person re-identification (re-ID) relies on discriminative features, which not only capture different spatial scales but also encapsulate an arbitrary combination of multiple scales.
Re-ID是一个示例级识别问题,它依赖于判别特征,这些特征不仅捕捉不同的空间尺度,还封装了多个尺度的任意组合。

We call features of both homogeneous and heterogeneous scales omni-scale features.
本文将均质和非均质尺度的特征称为omni-scale feature.(omni 全方位)。

本文设计一个新的re-ID网络,来捕获全方位特征,称该网络为OSNet。

模型在六个Re-ID数据集,Re-ID属性数据集,Cifa100,ImageNet上都表现出良好特性。

网络架构

(1)为了减少参数,设计一个Lite 3X3,如下图(b)所示。

论文阅读: Omni-Scale Feature Learning for Person Re-Identification
pointwise → depthwise instead of depthwise → pointwise
先使用了1x1的pointwise卷积,然后3x3 depth-wise卷积,作者说这样在re-Id上表现出更好的性能。

(2)Omni-Scale Residual Block
称图(b)为bottleneck.
论文阅读: Omni-Scale Feature Learning for Person Re-Identification
(3)网络基本结构
论文阅读: Omni-Scale Feature Learning for Person Re-Identification