Structure Aligning Discriminative Latent Embedding for Zero-Shot Learning.
Omkar Anil Gune, Biplab Banerjee, Subhasis Chaudhuri · DSpace (IIT Bombay) · 2018
We address the problem of zero-shot visual recognition in this paper and particularly focus on learning a discriminative latent embedding space where the visual image descriptors and the respective semantic class representations can be projected with coinciding alignment. While a supervised dimension reduction strategy which simultaneously optimizes the intra-class compactness and between-class separation is used to learn the latent space for the visual features, the semantic class prototypes are further projected onto this latent space via a multi-stage non-linear mapping function for re-alignment purposing. Furthermore, it is ensured that the visual and semantic class prototypes are likely to overlap in the latent space such that the overall similarity between samples from both the domains is maximized. Apart from remarkably reducing the so-called semantic gap, the discriminative property of the learned latent layer representations entails improved classification performance on both the standard zero-shot learning (ZSL) and the challenging generalized ZSL (GZSL) setups on three benchmark datasets (AWA, CUB, SUN) where the proposed method surpasses the state of the art results. © 2018. The copyright of this document resides with its authors.