Robust Partial-to-Partial Point Cloud Registration in a Full Range
Liang Pan, Zhongang Cai, Ziwei Liu
Abstract
Registration of 3D objects from point clouds is a challenging task due to sparse and noisy measurements, incomplete observations, and large transformations. In this work, we propose the <bold xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">G</b> raph <bold xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">M</b> atching <bold xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">C</b> onsensus <bold xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Net</b> work ( <bold xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">GMCNet</b> ) to estimate faithful correspondences for full-range Partial-to-Partial point cloud Registration (PPR) in object-level registration scenarios. To encode robust point descriptors, we employ a novel Transformation-robust Point Transformer ( <bold xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">TPT</b> ) module to adaptively aggregate local features with respect to the structural relations, taking advantage of both handcrafted rotation-invariant ( <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">RI</i> ) features and noise-resilient spatial coordinates. Based on the synergy of hierarchical graph networks and graphical modeling, we propose the Hierarchical Graphical Modeling ( <bold xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">HGM</b> ) architecture to encode robust descriptors comprising of i) a unary term learned from <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">RI</i> features, and ii) multiple smoothness terms encoded from neighboring point relations at different scales through our TPT modules. Extensive experiments show that GMCNet outperforms previous state-of-the-art methods for PPR.