In contrast to object recognition or object detection, which match data to existing object models, object discovery creates object models. Obviously, we need information sources to compensate for the lack of models. In this project, we investigate using 3-D motion of surface patches between multiple maps of the same environment as such a cue. Given multiple maps of a scene taken at different times, such that some foreground objects have moved between times at which we made maps, we produce a segmentation of each input 3-D scene into background and movable foreground objects. We also match object views segmented from scenes taken at different times (this clustering is technically the "object discovery" step).

Supplementary media for our ICRA2011 submission can be found here.