Point Cloud Registration with Self-supervised Feature Learning and Beam Search

Correspondence-free point cloud registration approaches have achieved notable performance improvement due to deep learning success, which optimizes the feature inference and registration in a joint framework. However, there are still several limitations that impede the effectiveness of practical applications. For one thing, most existing correspondences-free methods are locally optimal, and they tend to fail when the rotation is large. For another, when training a feature extractor, these approaches usually need supervised information from manually labeled data, which is tedious and labor-intensive. This paper proposes an effective point cloud registration method to resolve these issues, which is built upon a correspondence-free paradigm. Our approach combines self-supervised feature learning with a beam search scheme in the 3D rotation space, which can well adjust to the case of large rotation. We conduct extensive experiments to demonstrate that our approach can outperform state-of-the-art methods in terms of efficiency and accuracy across synthetic and real-world data.