講演情報
[7J-05]Efficient Hierarchical Structure Construction towards Correspondence Matching for Point Cloud
*Yu Ziyang1、常 穹1、宮崎 純1 (1. 東京科学大学 情報理工学院 情報工学系)
発表者区分：学生
論文種別：ロングペーパー
インタラクティブ発表：あり
論文種別：ロングペーパー
インタラクティブ発表：あり
キーワード：
Point Cloud Alignment、LiDAR Point Cloud、Information Retrieval
Point Cloud Alignment is a basic problem in 3D computer vision. Current alignment methods can be split into point-based and feature-based alignment methods. Compared to point-based methods, feature-based alignment methods can simultaneously capture global and local information for fine-grained point clouds. However, it is time-consuming to operate correspondence matching between high-dimension features constructed for further alignment. This research leverages HNSW, a method highly proficient in high-dimension nearest neighbor search, to enhance the efficiency of FPFH descriptor matching. Additionally, a Light-FPFH descriptor is proposed, which facilitates the optimization of HNSW implementation through the effective utilization of SIMD instructions, all while preserving matching accuracy. Experiments on fine-grained LiDAR point clouds demonstrate that our method provides x20-30 speed-up compared to tree-based alignment methods.