Pabsco: parameterized B-rep-based surface correspondence estimation for category-level 3D object matching applicable to multi-part items

In order to apply a 6DoF object pose and size estimation method to real-world scenarios such as robot picking in logistics warehouses, it is necessary to achieve both high recognition accuracy and low preparation costs. In this paper, we propose a new category-level 3D object matching method that can simultaneously address these two challenges. The proposed method estimates the 6DoF pose and size of unseen objects with unknown part sizes belonging to a target category by iteratively performing local matching between the surfaces of category-level 3D models represented by a combination of parameterized surfaces and the surfaces extracted from the scene. Additionally, it verifies the global consistency, such as occlusion-induced size estimation errors, by using a tree search. By leveraging these techniques, it becomes possible to efficiently and accurately estimate the 6DoF pose and size of multiple instances. We evaluated the proposed method through experiments simulating item picking in logistics warehouses and confirmed that it can estimate the 6DoF pose and size of logistics items with multiple parts at a recognition rate of over 92%, significantly higher than the 55% achieved by conventional methods.