3D pose estimation for symmetric and nonsymmetric objects

Abstract

Autonomous systems have to understand the 3D space, being able to detect objects and infer their pose to pick them and reliably perform a certain goal-oriented action. An increasing number of works focus on this topic motivated by self-driving cars or the Amazon picking challenge. In particular, we focus on pose estimation, a well-known problem in computer vision and robotics which is essential for object manipulation. This requires reliable identification of object poses to know how to pick them up in order to interact with them for a certain goal. Additionally, pose estimation involves several difficulties. Objects may have rotational symmetries or their appearance can vary significantly depending on the lighting or occlusions. A common approach to pose estimation is to first estimate a coarse pose to initialize ICP and get a fine pose estimation. We follow this idea in this work by comparing an object in an RGB-D setting to a set of views of the same CAD model obtained offline. Using Convolutional Neural Networks, we embed the images to a common space where they can be efficiently compared. Additionally, we propose to consider symmetries directly in the comparison to avoid inconsistencies in the pose estimation. Given the lack of benchmarks with symmetric objects for pose estimation, we obtain 6669 CAD models of very different kinds and generate realistic simulations of tabletop scenarios to train and test our approach. We also leverage a non-published dataset of real objects with symmetries. Finally, we infer rotational symmetries in new CAD models, obtaining a high recall and promising results that suggest further research.