Learning Local Reference Frame for 3D Point Cloud via Contrastive Learning

The local reference frame (LRF) plays a crucial role in 3D point cloud-based vision tasks. Most existing methods rely on manually designed rules. However, these handcrafted LRFs are sensitive to noise and exhibit poor repeatability and robustness. This paper proposes a coarse-to-fine learning framework that first establishes an initial LRF based on handcrafted rules and then refines the initial LRF using a rotation regression network. Furthermore, a contrastive learning-based training architecture is designed to implicitly supervise the learning of LRFs by aligning the source and target point clouds, eliminating the need for ground-truth LRF labels during training. Experimental results demonstrate that the proposed method achieves state-of-the-art repeatability and robustness on three widely used public datasets, particularly in the presence of interference such as noise, occlusion, and resolution variations.