Reference from: Heart of the machine
The article was published in early 2020. The author surveyed 245 papers for pedestrian re-identification (Person Re-identification) for the past two or three years, classified as closed-world ReID and open-world ReID, and reviewed the technical progress in this direction. The development gives several valuable directions, and is the most recent ReID review worth reading.
Information about the author of this article:
The following figure shows the five major steps of ReID technology summarized by the author:
1) Data collection;
2) Enclosure generation;
3) Training data annotation;
4) Model training;
5) Pedestrian retrieval
The author divides ReID technology into two major subsets: Closed-world and Open-world: it
can be seen that in terms of whether it is heterogeneous data, whether the annotation is complete, and whether it contains noise, etc., Open World ReID is closer to practical application.
Closed-world ReID technology
1) Learning method of feature representation:
2) Design of loss function in metric learning:
In addition, in the training strategy, data sampling with unbalanced samples should be considered.
3) Re-ranking optimization (re-ranking):
Statistics of data sets commonly used in closed world ReID:
Visualization of the accuracy of the SOTA method on the four data sets of the image-based ReID method: Visualization of the accuracy of the SOTA method on the four data sets of the
video-based ReID method
Open world ReID method
1) Heterogeneous data ReID
Based on deep ReID;
Text to image ReID;
Visible to infrared ReID;
Cross-resolution ReID;
2) End-to-end ReID
ReID of pure image / video;
ReID tracked by multiple cameras;
3) Semi-supervised and unsupervised ReID
Statistics of unsupervised ReID SOTA method:
4) Noise robust ReID
5) Open collection ReID
Outlook
The author proposes a new evaluation standard mINP to measure the quality of the ReID algorithm:
Based on the SOTA algorithm BagTricks, the AWG method is proposed:
The AGW method has achieved significant accuracy improvements on several large data sets:
The AGW method is very worthy of reference. The author expects it to become a strong baseline for future ReID research, and the code will be open source.
Thesis address:
https://arxiv.org/abs/2001.04193v1
AGW open source address:
https://github.com/mangye16/ReID-Survey
