Point cloud registration with rotation-invariant and dissimilarity-based salient descriptor

Point cloud registration is fundamental for generating a complete 3D scene from partial scenes obtained using 3D scanners. The key to the success of the task is extracting consistent descriptor values from arbitrary rotations (that is, rotation-robust) and selecting salient descriptors between the discriminative and noise regions. However, extracting rotation-invariant features and setting a criterion for selecting salient descriptors are difficult. In this study, we present a new method for building salient and rotation-robust 3D local descriptors for point registration tasks. For each point, we first extracted neighboring points (patches) for each interest point and aligned them with their local reference frame. Next, we encoded the aligned patches using cylindrical kernels to obtain rotation-invariant descriptors. Then, we estimated dissimilarities between the descriptors to exclude descriptors extracted from monotonous and repeating areas and select discriminative descriptors, and registration tasks were performed using only the salient descriptors. We experimented with our method on indoor and outdoor datasets (3DMatch and ETH) to evaluate its performance. The performance of our method is comparable to that of state-of-the-art methods using discriminative rotation-invariant descriptors. By selecting salient descriptors, we improved our feature-matching recall scores by up to 3%.