Interpretable feature maps for robot attention

Abstract

Attention is crucial for autonomous agents interacting with complex environments. In a real scenario, our expectations drive attention, as we look for crucial objects to complete our understanding of the scene. But most visual attention models to date are designed to drive attention in a bottom-up fashion, without context, and the features they use are not always suitable for driving top-down attention. In this paper, we present an attentional mechanism based on semantically meaningful, interpretable features. We show how to generate a low-level semantic representation of the scene in real time, which can be used to search for objects based on specific features such as colour, shape, orientation, speed, and texture.