Integrating Linear Skip-Attention With Transformer-Based Network of Multi-Level Features Extraction for Partial Point Cloud Registration

Abstract

Accurate point correspondences is critical for rigid point cloud registration in correspondence-based methods. Many previous learning-based methods employ encoder-decoder backbone for point feature extraction, while applying attention mechanism for sparse superpoints to deal with the partial overlap situation. However, few of these methods focus on the intermediate layers yet mainly pay attention on the top-most patch features, thus neglecting multi-faceted feature perspectives leading to potential overlap areas estimation inaccuracy. Meanwhile, obtaining correct correspondences is usually interfered with the one-to-many case and outliers. To address these issues, we propose a multi-level features extraction network with integrating linear dual attention mechanism into skip-connection stage of encoder-decoder backbone, both efficiently suppressing irrelevant information and guiding residual features to learn the common regions on which the network should focus to tackle the overlap estimation inaccuracy issue, combined with a parallel-structured decoder forming distinguishable features and potential overlapping regions. Additionally, a two-stage correspondences pruning process is designed to tackle the mismatch issue, which mainly depends on the rigid geometric constraint. Extensive experiments conducted on indoor and outdoor scene datasets demonstrate our method's accuracy and stability, by outperforming state-of-the-art methods on registration recall.