Graph Attention based Point Cloud Registration

Abstract

Point cloud registration is a fundamental but important technique in robotics and computer vision, such as 3D reconstruction, simultaneous localization and mapping (SLAM). The classical methods used hand-craft features extracted from each point cloud to estimate the rigid transformation between point clouds with iterative closest points (ICP) or its variants. Recently, deep learning has been widely used in object detection, segmentation and registration, especially the well-known work, PointNet, changed how we think about the representation of point clouds. However, the local information is ignored in PointNet as many works pointed out. In this paper, we advise to use graph attention to aggregate local features and a kind of cross attention method for point cloud registration. Firstly, we use a mini-ConvNet to extract point-wise features for sampled points and their neighbors. Then graph attention is applied to aggregate the local information from neighbor points to center points. We consider that previous works which directly estimate the transformation using the features from two point clouds ignore some relationship between the point clouds. Instead, we propose to explore the relation with a kind of cross attention. We perform extensive experiments to validate the effectiveness of our method.