Move and the Robot will Learn: Vision-based Autonomous Learning of Object Models

As robots are increasingly deployed in complex real-world domains, visual object recognition continues to be an open problem. Existing algorithms for learning and recognizing objects are predominantly computationally expensive, and require considerable training or domain knowledge. Our algorithm enables robots to use motion cues to identify and focus on a set of interesting objects, automatically extracting appearancebased and contextual cues from a small number of images to efficiently learn representative models of these objects. Learned models exploit complementary strengths of: (a) relative spatial arrangement of gradient features; (b) graph-based models of neighborhoods of gradient features; (c) parts-based models of image segments; (d) color distributions; and (e) mixture models of local context. The learned models are used in conjunction with an energy minimization algorithm and a generative model of information fusion for reliable and efficient recognition in novel scenes. The algorithm is evaluated on mobile robots in indoor and outdoor domains, and on images from benchmark datasets.