ATNet: Unsupervised Point Cloud Registration Network Integrating Adaptive Graph Convolution and Transformer

作者：Jiajin Du, Zhengyao Bai, Xuheng Liu, Zekai Li, Xiao Xiao, Yihan Zhang · 年份：2023 · DOI：10.1109/cac59555.2023.10452004 · 被引用次数：1 · 研究领域：Remote Sensing and LiDAR Applications、3D Surveying and Cultural Heritage、3D Shape Modeling and Analysis

Unsupervised point cloud registration algorithms often suffer from poor registration accuracy when dealing with partially overlapping point clouds due to the lack of effective internal monitoring. Traditional convolution methods, such as dynamic graph convolution, tend to extract partially indistinguishable features, resulting in limited ability to learn features. To address these challenges, inspired by adaptive graph convolution dynamically learning point cloud features based on adaptive kernels for point cloud generation, we propose an unsupervised point cloud registration network called ATNet, integrating adaptive graph convolution and transformer to improve point cloud evaluation by ensuring domain consistency. Firstly, we employ adaptive graph convolution to extract point cloud features, transformer to encode these features to preserve geometric information.Next, we use the graph optimization module to estimate a point-to-point graph by combining neighborhood matching scores. The aggregated neighborhood information is helpful to the generation of pseudo-target point clouds and high-quality correspondences. The point cloud evaluation module takes advantage of significant structural differences between outliers in the source and related pseudo-target clouds, while the structural differences of the interior points are minimal. By leveraging these geometric structural differences, the module scores the confidence of each estimated correspondence. Subsequently, the network u...