Pedestrian re identification (Reid) - Overview

What is? Re-ID?

• • 
    
• Pedestrian recognition （Person re-identification, abbreviation Re-ID） Also known as pedestrian recognition , Is the use of computer vision technology to determine whether there is a specific pedestrian in the image or video sequence . It is widely regarded as a sub problem of image retrieval . Given a monitored pedestrian image , Retrieve the pedestrian image under the cross device . It aims to make up for the visual limitations of the current fixed camera , And can detect with pedestrians / Combination of pedestrian tracking technology , It can be widely used in intelligent video surveillance 、 Intelligent security and other fields .
• As shown in the figure below ： There are multiple cameras shooting video sequences in an area ,ReID The requirements of a camera under the interest of pedestrians , Retrieve all pictures of the pedestrian under other cameras .
  • 
    

---

Why Re-ID？

In surveillance video , Due to the camera resolution and shooting angle , Very high quality face images are usually not available . When face recognition fails ,ReID It has become a very important alternative technology .

---

Research forms

• • 
    
• Data sets are usually pedestrian images obtained by manual annotation or detection algorithms , At present, it is independent of detection , Pay attention to identification
• Data sets are divided into training sets 、 Verification set 、Query、Gallery
• Train the model on the training set , After getting the model, right Query And Gallery Image feature extraction and similarity calculation in , For each Query stay Gallery Find out before N A similar picture
• Training 、 The identity of the person in the test is not repeated
  • 
    

---

Two directions

• • 
    
• feature extraction ： Learn to be able to cope with the characteristics of people changing under different cameras
• Measure learning ： Mapping the learned features to a new space makes the same people closer and different people farther
  • 
    

---

There are challenges

• • 
    
• Different cameras cause great changes in the appearance of pedestrians ;
• Target occlusion （Occlusion） Some features are lost ;
• Different View,Illumination Differences in characteristics that lead to the same goal ;
• Different target clothes have similar colors 、 Feature approximation leads to a decrease in discrimination ;
  • 
    

---

Common data set

CUHK03

Market1501

DukeMTMC-reID

MSMT17

Only commonly used data sets are listed here , A more complete data set can be referred to ：​ ​Person Re-identification Datasets​​

---

Commonly used evaluation index

• • 
    
• rank-k： The sorting list returned by the algorithm , front k If the bit is an existing search target, it is called rank-k hit .eg：rank1： The first is the search target rank-1 hit .
• Cumulative Match Characteristic (CMC)
  • 
    

Take a very simple example , Suppose in face recognition , There are 100 personal , Now comes 1 A face to be recognized （ If label by m1）, After comparing with the faces in the bottom database, the faces in the bottom database are sorted from high to low , We found that ：
If the recognition result is m1、m2、m3、m4、m5……, Now rank-1 The accuracy of is 100%;rank-2 The correct rate of is 100%;rank-5 The correct rate of is 100%;
If the recognition result is m2、m1、m3、m4、m5……, Now rank-1 The accuracy of is 0%;rank-2 The accuracy of is 100%;rank-5 The correct rate of is 100%;
If the recognition result is m2、m3、m4、m5、m1……, Now rank-1 The accuracy of is 0%;rank-2 The accuracy of is 0%;rank-5 The accuracy of is 100%;
Empathy , When there are many faces to be recognized , Take the average . For example, the face to be recognized has 3 individual （ If label by m1,m2,m3）, Similarly, there is a score from high to low for everyone's face ,
such as ：
Face 1 The result is m1、m2、m3、m4、m5……,
Face 2 The result is m2、m1、m3、m4、m5……,
Face 3 result m3、m1、m2、m4、m5……,
Now rank-1 The accuracy of is （1+1+1）/3=100%;
rank-2 The correct rate of is （1+1+1）/3=100%;
rank-5 The correct rate of is （1+1+1）/3=100%;
such as ：
Face 1 The result is m4、m2、m3、m5、m6……,
Face 2 The result is m1、m2、m3、m4、m5……,
Face 3 result m3、m1、m2、m4、m5……,
Now rank-1 The accuracy of is （0+0+1）/3=33.33%;
rank-2 The accuracy of is （0+1+1）/3=66.66%;
rank-5 The correct rate of is （0+1+1）/3=66.66%;

curve: Calculation rank-k Hit rate of , formation rank-acc The curve of , Here's the picture :

• mAP(mean average precision)： Reflect the extent to which all the correct pictures in the database of the person who searched are in front of the sorted list , It can be measured more comprehensively ReID Performance of the algorithm . Here's the picture , Suppose the search pedestrian is gallery There is 10 A picture , In the list Middle position （rank） Respectively 1、2、3、4、5、6、7、8、9, be ap by (1/ 1 + 2 / 2 + 3 / 3 + 4 / 4 + 5 / 5 + 6 / 6 + 7 / 7 + 8 / 8 + 9 / 9) / 10 = 0.90;ap large , The search results of this pedestrian are relatively high , For all query Of ap Take the average value to mAP
  

Generally speaking ,Precision It's the retrieved items （ such as ： file 、 Web page, etc ） How much is accurate ,Recall It's how many of the exact entries have been retrieved .
Accuracy rate = The number of positive samples detected / Total number detected
Recall rate = The number of positive samples detected / Number of all positive samples
Let's take a new example .
Suppose there is a search engine , According to search engines , The results are as follows :
Search for １ The total number of relevant samples is ５ individual : just , just , just , just , just

Rank１	just	negative	just	negative	negative	just	negative	negative	just	just
Recall	1/5=0.2	1/5=0.2	2/5=0.4	2/5=0.4	2/5=0.4	3/5=0.6	3/5=0.6	3/5=0.6	4/5=0.8	5/5=1.0
Precision	1/1=1.0	1/2=0.5	2/3=0.66	2/4=0.5	2/5=0.4	3/6=0.5	3/7=0.42	3/8=0.38	4/9=0.44	7/10=0.5

Precision From left to right 1/1, 1/2, 2/3, 2/4… And so on 　
Search for ２ There are a total of ３ individual , The following are the results returned by the search engine

Rank１	just	negative	negative	just	just	negative	negative
Recall	0.33	0.33	0.33	0.66	1	1	1
Precision	1.0	0.5	0.３3	０.5	0.６	0.5	0.43

We put each positive sample corresponding to Precision Averaging
Search for １ Of mAP:ｍAP = (1/1 + 2/3 + 3/6 + 4/9+ 5/10) / 5 = 0.72
Search for ２ Of mAP: ｍAP = (1/1 + 2/4 + 3/5) / 3 = 0.63
Holistic mAP = (0.72 + 0.63) /2 ＝ 0.675