Enhancing Mesh and Point Cloud Similarity Detection through Geometric Features and ICP

In this work, we propose a novel approach to assess the similarity between 3D data, aimed at enhancing the accuracy and robustness of 3D shape analysis. Our approach begins with the extraction of geometric features, including the dimensions of the oriented bounding box (OBB), Betti numbers indicating topological complexity, and variance metrics derived from surface normals and Fast Point Feature Histograms (FPFH). These features capture both the geometric and topological essence of the shapes, facilitating a nuanced preliminary similarity assessment. Following this, we employ a refined Iterative Closest Point (ICP) algorithm, enhanced with Random Sample Consensus (RANSAC), to meticulously align and compare the top 50 preliminary matched parts. This two-stage process leverages geometric feature normalization and Manhattan distances for initial part selection, before applying a PCA -oriented ICP for precise alignment. Our method addresses the common challenges in aligning parts with significant dimensional disparities, offering an improved framework for mesh and point cloud similarity detection. The experiments on a dataset containing 73 3D metal parts reveal that the proposed approach outputs the top choice of domain experts for 48 metal parts. Furthermore, when the top-5 similar parts are returned by the approach, in 66/73 of cases, the output contains the experts' choice.