| Περιγραφή: |
Person re-identification (re-ID) is an important technique that can be deployed in urban security surveillance systems, and it has attracted great attentions in the past decades. In this thesis, we study person re-identification with and without supervision. Firstly, in the study of person re-identification with supervision, we focus on person search. Person search aims to search for a target person among multiple images recorded by multiple surveillance cameras, which faces various challenges from both pedestrian detection and person re-identification. Besides the large intra-class variations owing to various illumination conditions, occlusions and varying poses, background clutters in the detected pedestrian bounding boxes further deteriorate the extracted features for each person, making them less discriminative. To tackle these problems, we develop a novel approach which guides the network with segmentation masks so that discriminative features invariant to the background clutters can be learned. We demonstrate that joint optimization of pedestrian detection, person re-identification and pedestrian segmentation enables to produce more discriminative features for pedestrian, and consequently leads to better person search performance. Secondly, in the study of person re-identification without supervision, we first study unsupervised domain adaptive person re-identification from two different views. From the first view, one effective way to tackle unsupervised domain adaptation (UDA) on person re-identification is to use clustering-based self-training approach, where a model is trained with hard pseudo-labels obtained from a clustering method. Using a hard pseudo-label, a sample is assigned to the cluster with the highest probability, which is sensitive to the incorrect clustering result due to imperfect clustering algorithms. Soft pseudo-labels can mitigate this issue by representing the sample with the full range of class probabilities from all clusters. Specifically, soft pseudo-labels comprise probabilities of full range classes, because they consider both the hard samples and easy samples. This will distract the model from learning more discriminative features in the hard examples. To solve this issue, we propose a coarse-to-fine refinement mechanism to produce robust refined soft pseudo-labels by progressively focusing more on the hard samples while less on the easy samples. The proposed refined soft pseudo-labels can be readily integrated into cross-entropy loss as a strong supervision to guide the model to learn more discriminative features. From the second view, data distribution alignment and clustering-based self-training are two feasible solutions to tackle unsupervised domain adaptation on person re-identification. Most existing alignment-based methods solely learn the source domain decision boundaries and align the data distribution of the target domain to the source domain, thus the re-ID performance on the target domain completely depends on the shared decision boundaries and how well the alignment is performed. However, two domains can hardly be precisely aligned because of the label space discrepancy of two domains, resulting in poor target domain re-ID performance. Although clustering-based self-training approaches could learn independent decision boundaries on the pseudo-labelled target domain data, they ignore both the accurate ID-related information of the labelled source domain data and the underlying relations between two domains. To fully exploit the source domain data to learn discriminative target domain ID-related features, in this thesis, we propose a joint framework for UDA on person re-ID that can learn domain-specific ID-related features, while feature distance distributions are aligned across two domains. Notably, the distribution alignment is conducted in our proposed homogeneous distance space, and re-arranged by the newly designed stair-stepping alignment (SSA) matcher. Finally, also in the study of person re-identification without supervision, we investigate the challenging fully unsupervised person re-identification. One effective way to address fully unsupervised person re-identification is to use a clustering-based contrastive learning approach. All the person images are split into several clusters, and the extracted person representation is then pulled closer to its cluster centroid and pushed away from others. However, the quality of the cluster centroid is significantly affected by the cluster impurity issue. Thus, on the premise that we cannot solve the cluster impurity issue, finding clusters that include an instance's wrongly-clustered ground-truth positive instances and further fixing the resulting incorrect supervision signal becomes a challenging problem. This thesis proposes a simple yet effective approach to address this problem. Our method can effectively locate the clusters that potentially include the wrongly-clustered positive instances, namely, plausible proxy. Then we introduce credibility to measure how much we should treat the mined plausible proxy centroid as a positive supervision signal rather than entirely negative. |