VRP-UDF: Toward Unbiased Learning of Unsigned Distance Functions From Multi-View Images With Volume Rendering Priors.

Area of Science:

• Computer Vision
• Computer Graphics
• Machine Learning

Background:

• Unsigned distance functions (UDFs) are crucial for representing open surfaces.
• Current methods for inferring UDFs use handcrafted differentiable renderers, which suffer from bias, sensitivity to outliers, and scalability issues.

Purpose of the Study:

• To develop a novel, data-driven differentiable renderer for more accurate UDF inference.
• To introduce learned volume rendering priors as a robust alternative to handcrafted methods.
• To enhance geometric detail and surface reconstruction in 3D scenes.

Main Methods:

• A neural network is pre-trained to render unsigned distances into depth images, creating volume rendering priors.
• Learned priors are generalized for UDF inference from RGB images using alpha blending.
• Auxiliary point sampling priors and novel sampling schemes improve accuracy near zero-level sets.
• A surface refiner integrates the volume rendering prior with Gaussian reconstruction methods.

Main Results:

• The learned volume rendering prior is unbiased, robust, scalable, and 3D-aware.
• The method demonstrates superior performance in UDF inference compared to state-of-the-art techniques.
• The volume rendering prior effectively enhances other neural implicit representations like signed distance functions and occupancy.

Conclusions:

• The proposed data-driven differentiable renderer and volume rendering priors offer a significant advancement in UDF inference.
• This approach provides a general strategy for improving various neural implicit representations and surface reconstruction tasks.
• The method achieves superior results on benchmarks and real-world scenes.