Abstract: The recognition and retrieval of 3D surfaces is crucial in a number of applications such as bin picking, Simultaneous Localisation and Mapping (SLAM), and autonomous vehicles. Effective solutions to this problem must be resistant to real-world adverse matching conditions such as clutter and occlusion, while simultaneously requiring low execution times. This talk describes a novel complete pipeline to address this problem. The pipeline allows the efficient retrieval of objects similar to a partial query surface. As part of this, the novel RICI and QUICCI local 3D shape descriptors are proposed. These are shown to be robust to clutter, compact, highly descriptive, and can be inexpensively computed on the GPU. Retrieval of nearest neighbour descriptors is enabled through a novel indexing strategy called the "Dissimilarity Tree", which through the use of a voting scheme allows the efficient identification of similar objects in a database.

Short-Bio: Bart is a Post-Doctoral fellow at the Norwegian University of Science and Technology, Department of Computer Science, where he also received the PhD in 2021. His research focuses on 3D object retrieval and recognition, with a special focus on efficient local 3D shape descriptors. His work has received a number of awards, including the "3DOR 2024 best paper award", "Norges Regnesentrals Masterpris", and the GRSI replicability stamp to a total 4 of his papers.

--
Comunicaciones DCC