Mutual Hypothesis Verification for 6D Pose Estimation of Natural Objects

Estimating the 6D pose of natural objects, such as vegetables and fruit, is a challenging problem due to the high variability of their shape. The shape variation limits the accuracy of previous pose estimation approaches because they assume that the training model and the object in the target scene have the exact same shape. To overcome this issue, we propose a novel framework that consists of a local and a global hypothesis generation pipeline with a mutual verification step. The new local descriptor is proposed to find critical parts of the natural object while the global estimator calculates object pose directly. To determine the best pose estimation result, a novel hypothesis verification step, Mutual Hypothesis Verification, is proposed. It interactively uses information from the local and the global pipelines. New hypotheses are generated by setting the initial pose using the global estimation and guiding an iterative closest point refinement using local shape correspondences. The confidence of a pose candidate is calculated by comparing with estimation results from both pipelines. We evaluate our framework with real fruit randomly piled in a box. The potential for estimating the pose of any natural object is proved by the experimental results that outperform global feature based approaches.