PEAL: Prior-embedded Explicit Attention Learning for Low-overlap Point Cloud Registration

Learning distinctive point-wise features is critical for low-overlap point cloud registration. Recently, it has achieved huge success in incorporating Transformer into point cloud feature representation, which usually adopts a self-attention module to learn intra-point-cloud features first, then utilizes a cross-attention module to perform feature exchange between input point clouds. The advantage of Transformer models mainly benefits from the use of self-attention to capture the global correlations in feature space. However, these global correlations may involve ambiguity for point cloud registration task, especially in indoor low-overlap scenarios, because the correlations with an extensive range of non-overlapping points may degrade the feature distinctiveness. To address this issue, we present PEAL, a Prior-embedded Explicit Attention Learning model. By incorporating prior knowledge into the learning process, the points are divided into two parts. One includes points lying in the putative overlapping region and the other includes points located in the putative nonoverlapping region. Then PEAL explicitly learns one-way attention with the putative overlapping points. This simplistic design attains surprising performance, significantly relieving the aforementioned feature ambiguity. Our method improves the Registration Recall by 6+% on the challenging 3DLoMatch benchmark and achieves state-of-the-art performance on Feature Matching Recall, Inlier Ratio, and Registration Recall on both 3DMatch and 3DLoMatch.