Learning Visual Matching From Small-Size Samples

ZHOU, JIAHUAN

doi:https://doi.org/10.21985/n2-6wm2-0c59

Work

Learning Visual Matching From Small-Size Samples

Public

Download PDF

Download All Files (.zip)

Visual matching is an important and fruitful research topic in computer vision area. Starting from the early face recognition, super-resolution, object tracking to the most recent person re-identification, cross-model retrieval, visual matching plays an important role as the core component in these tasks. The quality of visual matching directly and largely influence of the ultimate performance of these tasks. This dissertation concentrates on developing effective and efficient visual matching learning algorithms to facilitate the critical small-size sample challenges in visual matching, that only very few labeled positive, even only one sample is available for a particular instance. A specific human-centric visual matching task, image-based person re-identification, is adopted to evaluate our proposed works. The goal of person re-identification generally refers to evaluating the similarity of a probe image from an unknown identity against a set of gallery images with known identities. The gallery images may be obtained from different cameras at a different time. Person re-identification still remains a critical yet very challenging task in video surveillance due to the general difficulties of the large and complex variations in the visual appearances of a person under various views, poses, illumination and occlusion conditions. Besides the aforementioned difficulties, another critical issue that the very few labeled positive samples of one identity and severely imbalanced negative samples significantly constricts the quality of learning visual matching in person re-identification. This dissertation presents various effective and efficient techniques to address the critical small-size sample challenges in visual matching across images: a global metric learning algorithm based on a novel proposed similarity constraint, termed reference constraint, only needs few-shot positive samples for learning without any requirement of negative samples; an online local metric adaptation algorithm which is adoptable to any feature descriptors and any global metrics by using only one positive and extra unlabeled negative samples for metric learning; an extended online joint multi-metric learning method to learn multiple sharing-based joint Mahalanobis metrics for the given unlabeled data, no supervision label is requirement; and a two-stage hierarchical local metric adaptation algorithm to joint enhance the local discriminant of both unlabeled and labeled data. All the aforementioned methods aim to solve the severe small-size sample problem by relaxing the requirement of a large number of labeled positives for learning. Extensive experiments under different task setting on different datasets have validated the effectiveness and efficiency of the proposed approaches in the domain of image-based person re-identification.

Creator