Learning improved representations in encoder-decoder networks for point cloud registration via point interaction...

Area of Science:

• Computer Vision
• 3D Geometry Processing
• Machine Learning

Background:

• Point cloud registration aligns multiple 3D scans but struggles with missing data and viewpoint variations.
• Existing encoder-decoder architectures for point cloud processing can be enhanced to improve registration outcomes.

Purpose of the Study:

• To improve point cloud quality for more robust registration.
• To introduce novel modules that enhance learned point representations within encoder-decoder networks.

Main Methods:

• Proposed two plug-and-play modules: a point moving module for refining irregular point cloud positions and a point attention offset module for improving point matching likelihood.
• Integrated these modules into popular encoder-decoder networks to enhance implicit point representations.

Main Results:

• Achieved state-of-the-art results on 3DMatch (90.6% RR, 72.4% IR) and 3DLoMatch (69.3% RR, 43.5% IR) datasets.
• Demonstrated significant improvements in registration accuracy by enhancing point cloud quality through learned representations.

Conclusions:

• Improving point cloud quality via enhanced learned point representations is a beneficial approach for point cloud registration.
• The proposed modules offer a practical method to boost registration performance in scenarios with incomplete or varied 3D data.