特征提取
global feature
借助人体图像分割
Human Semantic Parsing for Person Re-identification CVPR2018
把人体图像分割成几个部分做attention
多尺寸
Multi-scale Deep Learning Architectures for Person Re-identification ICCV2017
手工设计多尺度的inception流
Attention
- Harmonious AttentionCNN(HA-CNN)CVPR2018
软注意力(channel+spatial)+硬注意力(指导local brach)
local feature
人体部分/区域特征的集合,解决对齐问题,身体部分由人体姿态构建或者粗略的垂直分割完成
Auxilary 辅助方法
需要额外的标注信息
- 语意属性标注
- view
View Confusion Feature Learning for Person Re-identification
ICCV2019
通过分类器得到view信息,然后通过分类器对抗损失、视觉feature损失、sift损失迫使提取出viewpoint invariant的feature
- camera
- data augumentation
-
位姿约束
guild 指导符合原来的分类 -
相机风格信息
camstyle -
将外观和结构分离生成图片
-
无监督域适应
SPGAN eSPGAN
Person Transfer GAN to Bridge Domain Gap for Person Re-Identification- PTGAN
Wei et al. [35] handle the domain
gap problem by proposing a Person Transfer Generative
Adversarial Network (PTGAN), transferring the knowledge
from one labeled source dataset to another unlabeled target
dataset.
An image-image domain adaptation method [118]
is developed with preserved self-similarity and domaindissimilarity, trained with a similarity preserving generative adversarial network (SPGAN).
A Hetero-Homogeneous Learning (HHL) method [215] simultaneously considers the
camera invariance with homogeneous learning (image pairs
from the same domain) and domain connectedness with
heterogeneous learning (cross-domain negative pairs).
- 网络结构
OmniScale Network (OSNet) 多尺度
Multi-Level Factorisation Net (MLFN) 多层级
度量函数
- identity loss 分类损失
- varification loss
区分这一对图片是否是同一个人
- triplet loss
优化方法
re-ranking 使用gallery-to-gallery 的相似性来优化初始序列。
待解决问题
Scalable Re-ID
- Lightweight Model. 更轻型的结构、模型蒸馏
- Resource Aware Re-ID.根据硬件配置适应性地调整模型 Deep Anytime ReID (DaRe)
Domain-Specific Architecture Design
多基于分类网络,为re-id网络提出新的结构
related work
- 多尺度网络 参数过多
1.另外加局部特征分支
- 使用多尺度的卷积流
OSnet
- 多层级
- 使用附加的属性信息
- 融合多层信息
- deep supervision