Zero-Shot Learning (ZSL) has gained its popularity recently owing to its promising characteristic that requires no training data to recognize new visual classes. One key technique is to transfer knowledge from the seen classes to the new unseen classes in an intermediate embedding space for both visual and textual modalities. Therefore, the construction of the embedding space is extremely important. Manifold embedding is able to well capture the intrinsic structure of the embedding space. To this end, with the assumption that the distribution of the semantic categories in the word vector space has an intrinsic manifold structure, this paper proposes a Manifold Embedding based ZSL (ME-ZSL) approach by formulating the manifold structure for the visual to textual embedding with the intra-class compactness, the inter-class separability, and the locality preservation. The linear, closed-form solution makes ME-ZSL efficient to compute. Extensive experiments on the popular AwA and CUB datasets validate the effectiveness of ME-ZSL.
展开▼
