Curious George: An Attentive Semantic Robotic Vision System
Abstract
Our robot, Curious George, has participated in the Semantic Robot Vision Challenge (SRVC): http://www.semantic-robot-vision-challenge.org/ a contest for robots performing a fully automated visual scavenger hunt within a realistic environment. The contest features completely autonomous acquisition of training data from the World Wide Web (WWW)(sources such as Google Images for example), that is no human labeled imagery is provided. Also, the visual categories chosen by the contest organizers are announced only at the time of the contest, and they are drawn from an unrestricted set within indoor objects, so contestants are challenged with perhaps the most general vision task amongst major contests held in the Computer Vision community. The evaluation phase of the contest requires robots to use the visual appearance models they have learned from the web data to explore a realistic contest setting which contains many of the target objects, and to recognize as many objects as possible. Our entry, Curious George, has won the SRVC for both of the two years in which it has been held.