Best Next-Viewpoint Recommendation by Selecting Minimum Pose Ambiguity for Category-Level Object Pose Estimation

抄録

Object manipulation is one of the essential tasks for a home helper robot, especially in helping a disabled person to complete everyday tasks. For handling various objects in a category, accurate pose estimation of the target objects is required. Since the pose of an object is often ambiguous from an observation, it is important to select a good nextviewpoint to make a better pose estimation. This paper introduces a metric of the object pose ambiguity based on the entropy of the pose estimation result. By using the metric, a best next-viewpoint recommendation method is proposed for accurate category-level object pose estimation. Evaluation is performed with synthetic object images of objects in five categories. It shows the proposed methods is applicable to various kind of object categories.