An Approximation to Context-Aware Size Modeling for Referring Expression Generation

Abstract

In this paper we describe a methodology for modeling context-dependent fuzzy size categories like "small" and "large". We consider in this work that the context is fixed by a collection of crisp size values, so that the relativity in the definition of the categories is related to the distances between sizes in the context. Modeling visual concepts like those related to size is a key point, for instance, in the generation of referring expressions (conjunctions of properties) identifying objects in a certain visual scene. Taking context into account in the fuzzy modeling process is crucial in order to get human-like results. We illustrate our approach with several examples, comparing the results with other usual approaches to size category modeling.