Gojcic Zan ; Zhou Caifa ; Wegner Jan D. ; Wieser Andreas

Abstract

We propose 3DSmoothNet, a full workflow to match 3D point clouds with asiamese deep learning architecture and fully convolutional layers using avoxelized smoothed density value (SDV) representation. The latter is computedper interest point and aligned to the local reference frame (LRF) to achieverotation invariance. Our compact, learned, rotation invariant 3D point clouddescriptor achieves 94.9% average recall on the 3DMatch benchmark data set,outperforming the state-of-the-art by more than 20 percent points with only 32output dimensions. This very low output dimension allows for near realtimecorrespondence search with 0.1 ms per feature point on a standard PC. Ourapproach is sensor- and sceneagnostic because of SDV, LRF and learning highlydescriptive features with fully convolutional layers. We show that 3DSmoothNettrained only on RGB-D indoor scenes of buildings achieves 79.0% average recallon laser scans of outdoor vegetation, more than double the performance of ourclosest, learning-based competitors. Code, data and pre-trained models areavailable online at https://github.com/zgojcic/3DSmoothNet.