Object segmentation and reconstruction from multi-view RGBD camera data

Object detection and recognition are usually prerequisites to most robot manipulation tasks. As such, a plenitude of recognition algorithms have been proposed, based both on visual appearance and shape cues.

In this thesis project you will concentrate on a commonly used sub-task in object recognition: namely, segmentation. Segmentation is the task of partitioning input data into classes, where all members of a class belong to the same physical object. You will implement / evaluate existing implementations of state of the art segmentation algorithms for RGBD data. You will investigate methods to fuse frame-based segmentation results in a volumetric model and instantiate object-centric object representations. You will then leverage labeled examples to learn a prior for reconstructing predictive models of the occluded portions of segmented objects.

The predictive models will be evaluated in the context of grasp planning for a simple parallel jaw gripper. 

Contact: Todor Stoyanov