Neural Implicit Representations for Multi-View Surface Reconstruction: A Survey.

Area of Science:

• Computer Vision
• 3D Geometry
• Machine Learning

Background:

• Conventional explicit geometric representations struggle with complex 3D shapes.
• Neural implicit representations offer a continuous, function-based approach to encoding 3D surfaces.
• Techniques include signed distance fields (SDF), unsigned distance fields (UDF), occupancy fields (OF), and neural radiance fields (NeRF).

Purpose of the Study:

• To systematically analyze the evolving methodologies and applications of neural implicit representations in 3D reconstruction.
• To provide a structured synthesis of research from 2020-2025.
• To guide emerging researchers and identify future research directions.

Main Methods:

• Categorization of neural implicit techniques based on geometric representation types (SDF, UDF, OF, NeRF).
• Classification of applications including object-level, scene-level, open-surface, and dynamic reconstruction.
• A dual-axis taxonomy to structure the survey of recent research (2020-2025).

Main Results:

• Neural implicit representations achieve superior multi-view reconstruction fidelity.
• They inherently support non-manifold geometries and complex topological variations.
• Demonstrated effectiveness across diverse reconstruction tasks.

Conclusions:

• Neural implicit representations are foundational tools in modern 3D reconstruction.
• The field is rapidly advancing, necessitating comprehensive analysis.
• This survey provides a roadmap for understanding current capabilities and future potential.