Abstract

In this paper we aim for zero-shot classification, thatis visual recognition of an unseen class by using knowl-edge transfer from known classes. Our main contributionis COSTA, which exploits co-occurrences of visual conceptsin images for knowledge transfer. These inter-dependenciesarise naturally between concepts, and are easy to obtainfrom existing annotations or web-search hit counts. We esti-mate a classifier for a new label, as a weighted combinationof related classes, using the co-occurrences to define theweight. We propose various metrics to leverage these co-occurrences, and a regression model for learning a weightfor each related class. We also show that our zero-shotclassifiers can serve as priors for few-shot learning. Exper-iments on three multi-labeled datasets reveal that our pro-posed zero-shot methods, are approaching and occasionallyoutperforming fully supervised SVMs. We conclude that co-occurrence statistics suffice for zero-shot classification.