TopFormer: Topology-Aware Transformer for Point Cloud Registration

Sheldon Fung, Wei Pan, Xiao Liu, John Yearwood, Richard Dazeley, Xuequan Lu

Abstract: The extraction of robust feature descriptors is crucial for achieving accurate point cloud registration. While the attention mechanism plays an important role in enabling sparse point features to learn global position-aware contextual information, the high sparsity at sub-sampled points can yield ambiguity in the corresponding features due to the loss of fine-grained structural information. In this paper, we propose TopFormer, a topology-aware Transformer that leverages surface-based geodesic topology to learn robust feature descriptors for point cloud registration. In particular, we design a topological structure encoding to capture point-pair surface-based structure in a sparse-through-dense manner. It couples the geodesic distance with the normal-based directional information, which provides a strong topological relation between each point pair. The proposed sparse-through-dense strategy is achieved by querying the information (e.g., geodesic distance) calculated from the dense point cloud for a pair of sparse points that exist in the dense point cloud. By doing so, the Transformer is able to learn feature descriptors that are more aware of the surface-based structural information. We evaluate the performance of our method on both indoor and outdoor datasets with different point cloud pair overlapping ratios. Experimental results show that our approach produces higher registration recalls than state-of-the-art techniques.