Dual-target Point Cloud Registration Using Representative Overlapping Points

Zhifeng Rao, Yuang Chen, Ming Cao, Zhiyun Lin

Keywords: point cloud registration, partially visible data, overlapping points, dual-target, alignment of attitude

Abstract:

Point cloud registration is a challenging task when only partially visible data is
available. Recently, many learning-based methods have been proposed for this
problem and have achieved satisfactory performance. However, they rarely combine multiple features and fail to pay attention to the key factor of registration:
alignment of attitude. Based on this phenomenon, we propose a dual-target point
cloud registration model, which combines multiple features learned from PointNet, DGCNN, and attention module. First, an initialization module is introduced
for coarse registration, in which a new target point cloud is added compared to
the original model. Second, we design a two-step attention-based representative
overlapping-point selection module, which uses both global and local features of
input point clouds. In the first step, overlapping scores are predicted using global
features. In the second step, a feature-matching matrix is obtained based on local
features and a self-attention module. Representative overlapping points are selected based on the overlapping scores in the first step and the feature-matching
matrix in the second step. Finally, a weighted SVD algorithm is used to estimate
the transformation from the point cloud after initialization to the target point cloud.
Extensive experiments on ModelNet40 show our method achieves state-of-the-art
performance compared to other learning-based methods. The code is available at
https://github.com/Dual-target.

Primary Area: applications to robotics, autonomy, planning

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 7056