DENSE THREE-DIMENSIONAL CORRESPONDENCE ESTIMATION WITH MULTI-LEVEL METRIC LEARNING AND HIERARCHICAL MATCHING

摘要

A method for estimating dense 3D geometric correspondences between two input point clouds by employing a 3D CNN architecture is presented. The method includes, during a training phase (100), transforming the two input point clouds (110, 120) into truncated distance function voxel grid representations (112, 122), feeding the truncated distance function voxel grid representations into individual feature extraction layers, extracting low- level features from a first feature extraction layer (140, 150), extracting high-level features from a second feature extraction layer (144, 154), normalizing the extracted low-level features and high-level features, and applying deep supervision of multiple contrastive losses and multiple hard negative mining modules (118, 118'). The method further includes, during a testing phase (200), employing high-level features capturing high-level semantic information to obtain coarse matching locations (240, 250), and refining coarse matching locations with the low-level features to capture low-level geometric information for estimating precise matching locations (210, 220).