Answers are grounded in this paper and cite their sources.
Zero-shot learning (ZSL) addresses the unseen class recognition problem by leveraging semantic information to transferknowledge from seen classes to unseen classes. Most existing models rely on learning an embedding space where theclasses of test samples according to sample-prototype similarities are determined. However, most approaches do not explorethe underlying distribution structure of embedded samples. In addition, whether the mapping can precisely preserve thesub-manifolds of samples is not investigated. In order to solve these issues, we proposed a transductive model named Graphand Low-rank based Cluster-Prototype Matching (GLCPM) model to learn a low-rank mapping to simultaneously explore thedistribution structure and keep the sub-manifolds of embedded samples. GLCPM aims to apply a teacher-student mode,where the embedding paradigm of the test semantic prototypes is guided by the training samples but keeps its own properties.Specifically, the learned visual embeddings of the test prototypes are encouraged to preserve the local intrinsic structure byconstructing a graph. Meanwhile, the visual features of test samples are clustered, and then a matching constraint is adoptedthat the projected test prototypes align the clusters so as to further improve the generality of the mapping function. Finally,an ensemble classifier, which combines the cluster-prototype similarity with the sample-prototype similarity, determines thelabel of each test sample. Extensive experiments on five benchmark data sets demonstrate the effectiveness of the proposedapproach.
No takes yet. Share an insight, caveat, or question.
Cao et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: