Press "Enter" to skip to content

Re-ID

本站内容均来自兴趣收集,如不慎侵害的您的相关权益,请留言告知,我们将尽快删除.谢谢.

什幺是行人重识别?

Person re-identification

Deep learning-based methods
两个基本模型下几个损失函数的精度

RGB image-based datasets
Video sequence datasets

future research directions
Multi-source多源数据行人重识别

写在前面的话

 

[1]Ye Mang,Shen Jianbing,Lin Gaojie,Xiang Tao,Shao Ling,Hoi Steven C H. Deep Learning for Person Re-identification: A Survey and Outlook.[J]. IEEE transactions on pattern analysis and machine intelligence,2021,PP.
[2]徐梦洋. 基于深度学习的行人再识别研究综述[C]//.中国计算机用户协会网络应用分会2018年第二十二届网络新技术与应用年会论文集.[出版者不详],2018:125-128.
[3]叶钰,王正,梁超,韩镇,陈军,胡瑞敏.多源数据行人重识别研究综述[J].自动化学报,2020,46(09):1869-1884.DOI:10.16383/j.aas.c190278.

 

很久没读英语了,如果笔记有语法错误别计较.

 

mark一下,道阻且长

 

Person re-identification

 

background

 

安防方面;用颜色直方图效果很差

 

Current Challenges:

 

 

Limited Training Data数据

 

Effectiveness性能

 

Efficiency效率

 

Domain Gap 能不能在realWord实用

 

 

实际场景中往往由于摄像头角度、穿着、光照以及天气的不同,导致模型的准度下降,考验模型的泛化能力,简单来说就是一般不能直接将训练好的模型直接应用到新场景、新dataset上,准确率低。

 

类内差异大于类间差异

 

 

    1. Uncontrained Environment(e.g.,Occlusion)实用

 

 

使用行人关键点检测,区分行人图片的遮挡部分和非遮挡部分,然后提取非遮挡部分的表征进行相似性匹配,忽略遮挡部分表征。But必须保证图像对齐性的高精准,否则将有可能出现身体部分错位的现象,增加模型的误判。

 

Potential Solutions:

 

 

Synthetic Pedestrain Images

 

Parts/Losses

 

Auto-ML/Pruning

 

Domain Adaptation

 

Alignment/3D Model

 

 

Deep learning-based methods

 

 

indentification deep model

regard the person ReID task as a classification issue

输出图像属性

architecture:

 

 

 

 

verification deep model

输入一对images,输出相似值来判断是否是同一个人

结合identification model和verification model has achieved promising results on person ReID

 

 

 

 

distance metric-based deep model

基于距离度量的深度模型的目的是使同一人物图像之间的距离尽可能小,而使不同人物图像之间的距离尽可能大

就是挖掘不同人物图像之间的相关性,并在训练阶段学习相似度度量

metric approach最常用三元组模型(triplet model)

 

 

 

 

part-based deep modelD

Attention mechanismas a part-feature learning module is adopted to enhance the discriminative ability of the learned deep feature.

The limitations :

a.The combination local and global feature is a good choice for person ReID.However,it may increase the complexity of model,thus reducing the training efficiency.

b.Ignore the pixel-level saliency

c. little consideration has been given to spatial contextual information

 

video-based deep model

深度学习技术也被应用到基于视频的人脸识别中

 

data augmentation deep model

Background:limited dataset can easily lead to overfitting

Conclusively, GAN-based data augmentation method could enhance the generalization capacity of ReID model.But the quality of generated images is relatively poor,thus bringing noise to the ReID system.

 

other perspective

camera network based methods

open-set person ReID

semi-supervised learning-based person ReID

low-resolution person ReID

Because the most frequently used sensor ——RGB sensor ,is sensible to lighting, occlusions and clutter conditions

开发了基于深度的和跨模态的模型:e.g. skeletal tracking、infrared、4D spatio-temporal signatures

 

 

两个基本模型下几个损失函数的精度

 

 

It ain’t hard to findidentification losses (i.e., Softmax and OIM) and distance metric-based losses(i.e.,Triplet and MSML) have complementary advantages to some extent.Thus, adopt jointly these two types of loss functions is a good choice for the CNN architecture.

 

Datasets

 

RGB image-based datasets

 

 

ViPER:star:️

背景、灯光条件、视角差别很大的

 

GRID

照明变化和低分辨率的

 

CUHK01

两个camera,一个正面或背面,另一个侧面的

 

CUHK03

the largest.

 

Market-1501

采集于清华大学校园中的6个不同视角的摄像头

the Deformable Part Model (DPM) detector captures the boxes of pedestrians自动检测并切割.

 

DukeMTMC-reID

来自8个高分辨率监控设备的1812个身份的36411个行人图像组成,其中1404个被两个以上的摄像头拍摄到,其余的被视为干扰识别

 

Airport

 

Partial-reID:star:️

60个行人的600张图像,每个行人有5张局部图像和5张全身图像

 

 

Video sequence datasets

 

 

3DPeS suffers from illumination and viewpoint variations. 光照和视角

 

ETHZ involves significant illumination variations and occlusions .光照和遮挡

 

PRID 2011 is captured in a relatively clean and simple scene and the dataset has consistent illumination changes. 干净

 

iLIDS-VID 视角、光照、服装相似、背景遮挡

 

MARS the largest

 

Multi-modal datasets

 

RGBD-ID is createdin different daysand the visual aspects of the pedestrians may change.

 

SYSU RGB-IR:star:️

 

由4台RGB cameras 和2台infrared cameras captured,红外和RGB模式之间存在巨大差异.

 

 

可以看到,deep learning稳定发挥,但是在几个dataset的表现上(i.e.,ViPER,Partial-ReID,SYSU RGB-IR)仍然有较大提高空间.

 

应该创建更接近真实场景的新数据集,来解决训练的模型在实际场景中鲁棒性较差的问题.

 

evaluation

 

 

    1. 累计匹配曲线 CMC :曲线表示一个查询目标出现在不同大小的候选表中的概率值,but始终只计算第一个被匹配的标注数据.

 

    1. Accordingly,存在多个标注时,采用 mAP 作为衡量标准.

 

 

future research directions

Future part-based deep model is advised to introduce relational information such as spatial context and temporal information to the local features.
忽略复杂背景,introduce the human body mask learning branch.
如何基于GAN生成高质量样本的数据集,设计GAN模型generate 视频序列的dataset,增强泛化能力.
person’s bbox大都假设框定好了但现实不是,handle the detection and re-identification jointly.
about datasets:Long term,Larger scale ,multi-modality data .
定义并优化位置的损失函数,并将其整合到最终的识别分数中,以减少检测的误差.

Multi-source多源数据行人重识别

 

1)不同的相机规格和设置,e.g.,high-resolution and low-resolution.

 

2)不同拍摄设备,非可见光设备e.g.,Infrared and Depth image.

 

3)文本信息.

 

4)由专家或者数字传感器自动获得的图像,素描与数字照片.

 

 

low-resulution 如何提高识别输入图像有效特征的准确性并尽可能少地引入与行人重识别无关或不利的视觉结果是提高低分辨率行人重识别的关键.

 

Infrared ReID 研究主要使用 特征空间投影转换 等方法解决跨模态特征匹配的问题, 但由于红外数据跨模态识别的独特之处在于照明类型的变化, 与完全依赖机器学习或基于不变特征提取的方法相比, 基于物理知识的 跨模态光度标准化建模 或许更有效.

 

Depth image 视角和距离都很大影响到深度信息的判别力.

 

文本 标记不完整但是准确性高

 

 

Multi-source data re-ID also need to solve the following problems:

Be First to Comment

发表回复

您的电子邮箱地址不会被公开。