Abstract
Vision based object grasping and manipulation in robotics require accurateestimation of object's 6D pose. The 6D pose estimation has received significantattention in computer vision community and multiple datasets and evaluationmetrics have been proposed. However, the existing metrics measure how well twogeometrical surfaces are aligned - ground truth vs. estimated pose - which doesnot directly measure how well a robot can perform the task with the givenestimate. In this work we propose a probabilistic metric that directly measuressuccess in robotic tasks. The evaluation metric is based on non-parametricprobability density that is estimated from samples of a real physical setup.During the pose evaluation stage the physical setup is not needed. Theevaluation metric is validated in controlled experiments and a new poseestimation dataset of industrial parts is introduced. The experimental resultswith the parts confirm that the proposed evaluation metric better reflects thetrue performance in robotics than the existing metrics.