3D Object Pose Estimation from RGB-D Data

Abstract

Estimation of position and rotation (pose) of a known object inside a scene is one of the fundamental tasks of computer vision. It has its use in robot picking in fully autonomous warehouses or augmented reality. We present a method for an object pose estimation, that utilizes scene depth and color information, but is able to function with depth alone. It is also capable of handling partial object occlusion and sensor noise. Resulting multi-platform C++ application uses both threads and OpenCL for parallelization, and was tested on synthetically generated scenes as well as scenes captured by depth cameras.