Synthesizing Object-Background Data for Large 3-d Datasets
David Breeden, Anuraag Chigurupati, Stephen Jay Gould, Andrew Y. Ng · 2008
In this project, we implement a 3-d data synthesis tool to generate a large number of object recognition training examples by synthesis in various configurations of relatively few background scenes and object models. Inputs for our tool are 2.5-d background scenes and 3-d object models. Each background scene is comprised of a visible-light image and corresponding depth map, collected by STAIR’s video camera and high-resolution laser scanner. We demonstrate a method to quickly and automatically generate scenes of the objects placed in the background scenes. These scenes are intended to be representative of scenes STAIR would encounter in the field containing real objects, and therefore good examples on which to train a STAIR classifier for object recognition. We compare the performance of an object recognition classifier trained on examples synthesized by our method against performance when trained on actual scenes.