1. Unsupervised Salience Learning for Person Re-identification
CVPR2013 Poster

2. Outline
Introduction
Method
Experiments
Conclusions

3. Introduction
Human eyes can recognize person identities based on some small salient regions.

4. Introduction
Person re-identification handles pedestrian matching and ranking across non-overlapping camera views.
viewpoint change and pose variation cause uncontrolled misalignment between images.

6. Introduction Motivations :
1. We can recognize persons across camera views from their local distinctive regions
2. Human salience
3. Distinct patches are considered as salient only when they are matched and distinct in both camera views

7. Method Dense Correspondence Unsupervised Salience Learning
Matching for Re-identification

8. Dense Correspondence Features: Adjacency constrained search:
dense color histogram + dense SIFT
Adjacency constrained search:
simple patch matching

9. Adjacency constrained search
Search set :

10. Adjacency constrained search
Adjacency Searching:

11. Unsupervised Salience Learning
two methods for learning human salience:
K-Nearest Neighbor Salience (KNN)
One-Class SVM Salience (OCSVM)

12. Unsupervised Salience Learning
Definition: Salient regions are discriminative in making a person standing out from their companions, and reliable in finding the same person across camera views.
Assumption: fewer than half of the persons in a reference set share similar appearance if a region is salient. Hence, we set k = Nr/2. Nr is the number of images in reference set.

13. Unsupervised Salience Learning

14. K-Nearest Neighbor Salience

15. K-Nearest Neighbor Salience

16. Unsupervised Salience Learning

17. One-Class SVM Salience

18. Matching for Re-identification
Bi-directional Weighted Matching
Complementary Comination

19. Matching for Re-identification

20. Experiments
Dataset :
VIPeR Dataset
ETHZ Dataset

21. Experiments

22. Experiments

23. Experiments

24. Experiments

25. Experiments

26. Conclusions
1. An unsupervised framework to extract distinctive features for person re-identification.
2. Patch matching is utilized with adjacency constraint for handling the misalignment problem caused by viewpoint change and pose variation.
3. Human salience is learned in an unsupervised way.