SuperUDF: Self-Supervised UDF Estimation for Surface Reconstruction.

Area of Science:

• Computer Vision
• Computer Graphics
• Machine Learning

Background:

• Surface reconstruction is crucial for 3D modeling and analysis.
• Unsigned distance functions (UDFs) offer advantages in representing open surfaces.
• Existing learning-based methods face challenges with sparse data and efficiency.

Purpose of the Study:

• To develop a self-supervised learning framework for UDF-based surface reconstruction.
• To improve the efficiency and robustness of UDF estimation, particularly for sparse sampling scenarios.
• To introduce a novel mesh extraction method from learned UDFs.

Main Methods:

• Proposed SuperUDF, a self-supervised learning approach for UDF estimation.
• Incorporated a learned geometry prior for efficient training.
• Introduced a novel regularization loss for robustness against sparse sampling.
• Developed a learning-based mesh extraction technique.

Main Results:

• SuperUDF demonstrates superior performance compared to state-of-the-art methods.
• Achieved improvements in both reconstruction quality and computational efficiency.
• Validated on multiple public datasets, showcasing robustness to sparse sampling.

Conclusions:

• SuperUDF provides an effective and efficient solution for learning-based surface reconstruction using UDFs.
• The proposed method advances the state-of-the-art in handling sparse data for 3D surface estimation.
• Future work includes releasing the code for broader research adoption.