Zero-shot Object Classification with Large-scale Knowledge Graph

Abstract

Zero-shot learning is research for predicting unseen categories, and can solve problems such as dealing with unseen categories that were not anticipated at the time of training and the lack of labeled datasets. One of the methods for zero-shot object classification is using a knowledge graph, which is a set of explicit knowledge. Since recognition is limited to the categories contained in the knowledge graph and the relationships among categories are expected to be quantitatively and qualitatively richer depending on the graph size, it is desirable to handle a large-scale knowledge graph that contains as many categories as possible. We use a knowledge graph that contains about seven times as many categories as the knowledge graphs used mainly in existing research to enable classification of a larger number of categories and to achieve more accurate recognition. When using large-scale knowledge graph, it is expected that the number of noisy nodes and edges will increase. Therefore we propose a method to extract useful information from entire graph by positional relationships between categories and the types of edges in the knowledge graph. We classify images that were unclassifiable in existing research and show that the proposed data extraction method improves performance compared to using entire graph.