Bootstrapping Visual Memories for a Humanoid Robot

This paper addresses the problem of creating visual memories of salient objects arising in a certain unknown environment and the ability of recalling them at a later time. We adopt a developmental approach, i.e., we start with low complexity and coarse information that is incrementally refined to create a hierarchy of visual representations for the observed scenes. Colour histogram-based measurements provide a simple and quick methodology to bootstrap the creation of visual memories. Experiments performed on the humanoid platform iCub illustrate how this technique can successfully create concepts by grouping related images together in clusters, subject to a similarity threshold.