FINet: Dual Branches Feature Interaction for Partial-to-Partial Point Cloud Registration

Data association is important in the point cloud registration. In this work, we propose to solve partial-to-partial registration from a new perspective, by introducing multi-level feature interactions between source and reference clouds at extraction stage, such that can be realized without attentions or explicit mask estimation for overlapping detection as adopted previously. Specifically, present FINet, interactionbased structure with capability enable strengthen information associating inputs multiple stages. To achieve this, first split features into two components, one rotation translation, based on fact they belong different solution spaces, yielding dual branches structure. Second, insert several interaction modules extractor data association. Third, transformation sensitivity loss obtain rotation-attentive translation-attentive features. Experiments demonstrate our method performs higher precision robustness compared state-of-the-art traditional learning-based methods. Code available https://github.com/megvii-research/FINet.