PointNorm-Net: Self-Supervised Normal Prediction of 3D Point Clouds via Multi-Modal Distribution Estimation.

Area of Science:

• Computer Vision
• 3D Geometry Processing
• Machine Learning

Background:

• Supervised deep normal estimators excel on synthetic data but fail in real-world scenarios due to domain gaps.
• Creating annotated real-world 3D data for normal estimation is costly and time-consuming.

Purpose of the Study:

• To develop the first self-supervised deep learning framework, PointNorm-Net, for accurate 3D surface normal estimation.
• To address the challenge of domain gap and data annotation costs in real-world 3D scenes.

Main Methods:

• Introduced PointNorm-Net, a self-supervised deep learning framework.
• Developed a three-stage multi-modal normal distribution estimation paradigm.
• The paradigm is adaptable to both deep and traditional optimization-based normal estimation.

Main Results:

• PointNorm-Net demonstrates superior generalization capabilities on real-world datasets.
• The proposed method outperforms existing conventional and deep learning approaches.
• Achieved state-of-the-art performance across three diverse real-world 3D datasets.

Conclusions:

• Self-supervised learning is a viable and effective approach for 3D surface normal estimation in real-world scenarios.
• PointNorm-Net offers a robust solution to overcome domain gap issues and reduce annotation dependency.
• The framework provides a significant advancement for 3D reconstruction and scene understanding tasks.