.png){kind=icon}
.png){kind=icon}
Return
Neural Implicit Representations for Multi-View Surface Reconstruction: A Survey
.png){kind=icon}
PRE
AI
DOI：10.1109/TVCG.2025.3582627.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Diverging from conventional explicit geometric representations, neural implicit representations utilize continuous function approximators to encode 3D surfaces through parametric formulations including signed distance fields (SDF), unsigned distance fields (UDF), occupancy fields (OF), and neural radiance fields (NeRF). These approaches demonstrate superior multi-view reconstruction fidelity by inherently supporting non-manifold geometries and complex topological variations, establishing themselves as foundational tools in 3D reconstruction. Neural implicit representations can be applied to a diverse array of reconstruction tasks, including object-level reconstruction, scene-level reconstruction, open-surface reconstruction and dynamic reconstruction. The exponential advancement of neural implicit representations in 3D reconstruction necessitates systematic analysis of their evolving methodologies and applications. This survey presents a structured synthesis of cutting-edge research from 2020-2025, establishing a dual-axis taxonomy that categorizes techniques by geometric representation types and application scenarios. Through this survey, we aim to familiarize emerging researchers with the current landscape of neural implicit representation in surface reconstruction, assess innovative contributions and limitations in existing research, and encourage prospective research directions.
Keywords:
3D surface reconstruction
neural implicit representation
volume rendering
multi-view stereo(MVS)
object-level reconstruction
scene-level reconstruction
open-surface reconstruction
dynamic reconstruction
Journal
IF:
6.5
Papers:
309
Citations:
2.2W