Person re-identification, also known as person re-identification, is widely considered to be an image retrieval sub-problem. It is a technology that uses computer vision technology to determine whether there is a specific pedestrian in an image or video, that is, given a monitored pedestrian image retrieval This pedestrian image across devices. Pedestrian re-identification technology can make up for the visual limitations of current fixed cameras, and can be combined with pedestrian detection and pedestrian tracking technology to be used in video surveillance, intelligent security and other fields.
Generally, pedestrian re-identification has a short-term effect. The clothes of the pedestrian we need to identify are a main feature. Of course, clothes are only one of the features. If the pedestrian changes clothes, the pedestrian re-identification may fail.
Its main principles are as follows
First we need a pedestrian detection system to detect all pedestrians in the original video frame to form a warehouse Gallery. Then we need a personal image of the pedestrian to be retrieved, called Probe, to extract the characteristics of the pedestrian and put it in the Gallery. Comparison is performed in , that is, the similarity is measured. If it is greater than a certain threshold, it is considered that the pedestrian has been retrieved in Gallery.
data set
The data set is usually pedestrian images obtained through manual annotation or detection algorithms. It is currently independent of detection and focuses on recognition. It is divided into training set, verification set, Query (a bunch of Probes, personal photos to be retrieved), and Gallery (image library). Carry out model training on the training set. After obtaining the model parameters, extract the feature values of the pictures in the Query and Gallery to calculate the similarity. For each Query, find the first N similar pictures in the Gallery. Character identities cannot be repeated during training and testing .
The data set is divided into single frame and sequence.
single frame
sequence
challenge
The current accuracy of pedestrian re-identification can only reach 90%. Different face recognition can reach an accuracy of 99%. The main reason is that
Commonly used evaluation indicators
- Rank-k: In the sorted list returned by the algorithm, if the search target exists in the top k positions, it is called a rank-k hit.
In the first row of the picture above, the first sample is hit, so it is a rank-1 hit; the second sample in the second row is not hit in the first 3, but the fourth one is hit. , it belongs to the top 5 hits, including rank-1 there, so it is a rank-5 hit; both the fourth and fifth rows do not have the top five hits, but are in the top 10 episodes, so they are both good rank-10. In this way, the probability of rank-1 is 40%, the probability of rank-5 is 60%, and the probability of rank-10 is 100%.
- CMC curve: Calculate the hit rate of rank-k and form the curve of rank-acc
The above picture is the rank-k curve of each algorithm. No matter which algorithm is used, the larger the k value, the higher the accuracy. This is inevitable.
- mAP curve: reflects the extent to which all correct pictures of the retrieved person are at the front of the sorted list in the database, and can more comprehensively measure the performance of the ReID algorithm.
In the above picture, if Gallery has multiple Probe1 pictures, one picture is hit at rank-1, and the other two pictures are hit at rank-4 and rank-9 respectively. Probe2 was hit at rank-2, rank-5, and rank-7. Since Probe1 has rank-1 and Probe2 only has rank-2, does it mean that Prob1 is better than Prob2? The answer is not necessarily.
Here we need to calculate the AP value. In Probe1, the probability of rank-1 is 1/1, the probability of rank-4 is 2/4 (rank-1 is included here), and the probability of rank-9 is 3/9 (here Including rank-1 and rank-4), adding the probabilities of rank-1, rank-4, and rank-9 and dividing it by the average of 3, we get the AP value of Probe1 as 61.1%. Similarly, the AP value of Probe2 is 44.3%.
mAP is the average of all Probe APs. Since there are only two Probes here, mAP=52.7%.
IntelliJ IDEA 2023.3 & JetBrains Family Bucket annual major version update new concept "defensive programming": make yourself a stable job GitHub.com runs more than 1,200 MySQL hosts, how to seamlessly upgrade to 8.0? Stephen Chow's Web3 team will launch an independent App next month. Will Firefox be eliminated? Visual Studio Code 1.85 released, floating window Yu Chengdong: Huawei will launch disruptive products next year and rewrite the history of the industry. The US CISA recommends abandoning C/C++ to eliminate memory security vulnerabilities. TIOBE December: C# is expected to become the programming language of the year. A paper written by Lei Jun 30 years ago : "Principle and Design of Computer Virus Determination Expert System"