IMRA's Contribution to Semantic Robot Vision Challenge

Semantic Robot Vision Challenge

The Semantic Robot Vision Challenge seeks to fuse the state of the art research from various vision and mobile robotics communities. In this challenge, robots must perform a scavenger hunt in a previously-unknown indoor environment. The robots will be given an electronic text file that will contain a list (in English) of objects that must be located in the environment. Before entering the competition arena, the robots will be given time to access the Internet and search it for examples of the objects in order to build a classification database. Once the robot has its database, it will be required to explore the competition arena and perform a computer vision-based search for these objects. At the end of the event, the robots are required to generate a set of images of the objects (taken from the robot's on-board cameras) where the locations of the objects are highlighted with a bounding box and labeled appropriately. Teams will be scored by how many objects have been correctly located.

Our Robot in SRVC08

In 2008 SRVC competition IMRA took the second place in robot league with 7 points (12 points for the first place).

- IMRA's robot was the only to compete in SRVC08 using fixed camera-based system for both object recognition and navigation (other teams used laser for navigation and some PTZ cameras).

- Only IMRA's detection results were fast i.e. provided < 10 sec after seeing an objects (others - up to several minutes) = "real" real-time bonus points.

- Only IMRA was using context to enforce recognition - e.g. "Spam" was found on a table.

Benchmarks

Robot League (recognition + navigation) was evaluated during SRVC08 competition and Software League (recognition only) version (download source code) was tested later with official 2007/2008 datasets. Results can be summarized as follows:

2008: Robot mode (navigation+recognition) - 20% total, 40% specific, 0% generic

2008: Software mode (recognition) - 50% total, 90% specific, 10% generic

2007: Software mode (recognition) - 57% total, 100% specific, 10% generic