Occluded person re-identification

Abstract

Person Re-identification aims at recognizing an individual who appears under different surveillance camera perspectives. With the development of deep neural networks, it has gained increasing interest in the computer vision community. However, the research in a real-world setting is more complicated. One important problem is that person images are often occluded by either an object (e.g. car) or another person. To solve the problem, a new task called Occluded Re-identification (ReID) is drawing increasing attention. This dissertation first examines existing Occluded ReID methods and reproduces several state-of-the-art methods. By analyzing the advantages of existing Occluded ReID methods, we design a powerful OccludedReID baseline, which can achieve state-of-the-art or satisfactory performance on three occluded/partial datasets. Meanwhile, we introduce some new settings to change the training domain of existing methods and obtain 87.3% rank-1 accuracy on the OccludedREID dataset, which is at least 5.7% better than existing state-of-the-art methods. Finally, some important yet under-investigated problems of existing methods are discussed.