Generalized Zero-shot Learning using Open Set Recognition

Omkar Gune (Indian Institute of Technology Bombay), Amit More (Indian Institute of Technology Bombay), Biplab Banerjee (Indian Institute of Technology Bombay), Subhasis Chaudhuri (Indian Institute of Technology Bombay)

Abstract
Generalized Zero-shot Learning (GZSL) aims at identifying the test samples which can belong to previously \textit{seen} (training) or \textit{unseen} visual categories by leveraging the side information present in the form of class semantics. In general, GZSL is a difficult problem in comparison to the standard Zero-shot Learning (ZSL) given the model bias towards the seen classes. In this paper, we follow an intuitive approach to solve the GZSL problem by adhering ideas from the Open Set Recognition (OSR) literature. To this end, the proposed model acts as a pre-processing module for the GZSL inference stage which decides whether a given test sample belongs to seen or unseen class (domain). In order to comprehend the same, we generate \textit{pseudo unseen visual} samples from the available seen data and further train a domain classifier for on-the-fly domain label assignment for the test samples. The domain specific inference modules are then applied subsequently for improved classification. We experiment on standard benchmark AWA1, APY, FLO, and CUB datasets which confirm superior performance over the existing state of the art.

DOI
10.5244/C.33.4
https://dx.doi.org/10.5244/C.33.4

Files
Paper (PDF)
Supplementary material (PDF)

BibTeX