Siamese-PointNet++: Point Cloud Classification with Siamese PointNet++

Abstract

It is significant to explore the related information of point pairs to improve the classification accuracy of a point cloud. This paper proposes the Siamese PointNet++, which is end-to-end trained offline with point-set pair. More specifically, PointNet++is used to extract features from point-set pairs, and then a relation module with 1D CNN architecture is applied to compute the relation scores. We conducted extensive experiments on the test data from the 3D point cloud classification challenge of the 2019 IEEE GRSS Data Fusion Contest. The inspiring experimental results demonstrate the effectiveness of the proposed framework.