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

Abstract

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

Geometric Transformer for Fast and Robust Point Cloud Registration

“…The first class [12], [23], [24], [25], [26], [27], [28] follows the idea of ICP [29], which iteratively establishes soft correspondences and computes the transformation with differentiable weighted SVD. And the second class [30], [31], [32], [33] first extracts a global feature vector for each point cloud and regresses the transformation with the global feature vectors. Due to the lack of a robust estimator, direct registration methods opt to adopt an iterative registration scheme to progressively refine the estimated transformation.…”

Scale robust point matching‐Net: End‐to‐end scale point matching using Lie group

“…Finally, the updated transformation h can be obtained by substituting a from Equations ( 11)- (16). The linearity of Lie algebra ensures that the elements of the next iteration are still in the Lie group, which means our method is structure preserving.…”

“…We compare SRPM-Net with two kinds of methods: (1) the traditional methods, such as ICP [1], Scale-ICP (SICP) [18] and RPM [7]; and (2) the deep learning methods, such as Deep Closest Point (DCP) [24], PointNetLK [10], RPMNet [14], RPMNet_corr [15] and FINet [16].…”

“…RPM‐Net_corr [15] defines the correspondence error as the metric using the incorrectly matched point pairs, and adopts the correspondence‐based loss function. The Feature Interaction‐based network (FINet) [16] introduces multi‐level feature interactions between the source and the reference clouds at the feature extraction stage, and estimates the transformation iteratively. These methods, however, only deal with the rigid transformation.…”