Robust and Fast Hypothesis Verification in 3D Object Recognition

3D object recognition requires generating and verifying hypothesis, where a hypothesis means a hypothesized object together with a hypothesized pose of the object in the scene. Hypothesis verification is typically achieved by transforming the hypothesized object model into the scene based on its pose, estimating the overlap between the transformed model and the scene, and accepting the hypothesis if the overlap area is sufficiently large. This paper develops a robust and fast hypothesis verification approach to improve the recognition accuracy without slowing down its speed. Our key idea is to identify corresponding points in the overlap between a hypothesized model and the scene using both point position and normal, and to speed up the identification of corresponding points by utilizing a signed distance transform map defined on the model’s voxelization. Moreover, we estimate overlap area from the identified corresponding points more accurately based on an idea of area-weighted vertices. Experiments on publicly available datasets demonstrate the robustness and efficiency of our verification approach.