Cross-Scenario Transfer Person Reidentification

Abstract

Person re-identification is to match images of the same person captured in disjoint camera views and at different time. In order to obtain a reliable similarity measurement between images, manually annotating a large amount of pairwise cross-camera-view person images is deemed necessary. However, such a kind of annotation is both costly and impractical for efficiently deploying a re-identification system to a completely new scenario, a new setting of non-overlapping camera views between which person images are to be matched. To solve this problem, we consider utilizing other existing person images captured in other scenarios to help the re-identification system in a target (new) scenario, provided that a few samples are captured under the new scenario. More specifically, we tackle this problem by jointly learning the similarity measurements for re-identification in different scenarios in an asymmetric way. To model the joint learning, we consider that the re-identification models share certain component across tasks. A distinct consideration in our multi-task modeling is to extract the discriminant shared component that reduces the cross-task data overlap in the shared latent space during the joint learning, so as to enhance the target inter-class separation in the shared latent space. For this purpose, we propose to maximize the cross-task data discrepancy (CTDD) on the shared component during asymmetric multi-task learning, along with maximizing the local inter-class variation and minimizing local intra-class variation on all tasks. We call our proposed method the constrained asymmetric multi-task discriminant component analysis (cAMT-DCA). We show that cAMT-DCA can be solved by a simple eigen-decomposition with a closed form, getting rid of any iterative learning used in most conventional multi-task learning. The experimental results show that the proposed transfer model gains a clear improvement against the related non-transfer and general multi-task person re-identification models.

Color Distribution Pattern Metric for Person Reidentification

“…Hu et al [7] employed deep learning in a transfer metric, which learned hierarchical nonlinear transformations of cross-domains by mapping discriminative knowledge between a labeled source domain and unlabeled target domain. Others [12] jointly learned a transfer metric in an asymmetric way by extracting discriminant shared components through multitask modeling to enhance target interclass differences under shared latent space. Recently, Shi et al [13] showed interest in attribute features extracted from semantic level for the Re-ID task and employed it in crossdomains transfer metric learning.…”

“…However, methods based on dense block features [5][6][7][8][9][10][11][12][13][14]16], which encode not only person pixel information but also background pixel information, cannot properly filter out background information due to pose variations and different body types. Furthermore, dense block features introduce distortion caused by resizing from different resolutions.…”

“…Figure 1 shows different body types (including massive build, medium build, and slim build), pose variations, and viewpoint variations (note that viewpoint variations in this paper refer to front view, back view, left view, and right view or any state in-between but do not include top view). Matching accuracies of these aforementioned methods [5][6][7][8][9][10][11][12][13][14]16] suffer from encoded background information due to inappropriate feature selection, in which case the fixed classification metric could not properly separate encoded person information from encoded background information, under significant pose variations and different body types. We refer to this kind of interference as background interference.…”

“…Furthermore, resolutions of pedestrian images are not required to be fixed in , different from fixed image resolutions in [5][6][7][8][9][10][11][12][13][14]16]. This characteristic of avoids distortion interference brought by resizing.…”

“…In our experiment, we loosely crop out background areas, as shown in Figure 4, instead of scaling pedestrian images into a fixed resolution like [5][6][7]12].…”