High density 3D correspondence estimation using multilevel metric learning and hierarchical matching

摘要

We propose a method for estimating three ‐ dimensional geometric correspondence between two input point clouds using a 3D CNN architecture.This method isDuring a training stageConvert two input point clouds to a cutoff distance function voxel grid representationThe cutoff distance function voxel grid representation is supplied to the individual feature extraction layer with the coupling weights;Extracting low level features from the first feature extraction layerExtracting high level features from the second feature extraction layerNormalized low level features and high level features are normalized to obtain unit vector features.Deep super vision is applied to a plurality of control losses and multiple hard negative mining modules of the first and second feature extraction layers.In addition, the method uses high level features to capture high level semantic information during the test stage and refines the coarse matching position with low level features to capture the coarse matching position and capture low level geometric information for estimating the precision matching position.Diagram