Abstract
Standard 3D reconstruction methods often fail to preserve sharp geometric features, resulting in aliased or rounded edges. We present PRS (Sharp Feature Priors), a data-driven framework that recovers these features by learning local surface-based fields. By predicting distance-to-feature, direction-to-feature, and surface improvement metrics, PRS enables an iterative remeshing process that achieves state-of-the-art accuracy across arbitrary resolutions without manual parameter tuning.
Introduction
In 3D computer vision and CAD, shapes are frequently represented as implicit functions (e.g., Signed Distance Fields) and converted into explicit meshes via algorithms like Marching Cubes. These methods typically struggle with sharp features, producing jagged approximations instead of clean 90-degree edges. While recent neural methods attempt to mitigate this, they are often "resolution-locked" to specific grid sizes. PRS addresses these challenges by providing a resolution-free, automatic pipeline to detect and reconstruct sharp features directly on the mesh.
Methodology
The core of PRS lies in three learned surface-based priors predicted on local triangulated patches:
- Distance-to-feature ($d$): A scalar field estimating the distance from a vertex to the nearest sharp edge.
- Direction-to-feature ($r$): A vector field pointing toward that edge.
- Surface improvement ($s$): An edge-based field predicting normal alignment improvement following an edge flip.
The PRS Pipeline
The refinement process follows a structured sequence:
- Feature Detection: A Dynamic Graph CNN (DGCNN) predicts the vertex-based $d$ and $r$ fields.
- Vertex Snapping: Vertices near features are "snapped" onto estimated sharp edges using the predicted distance and direction.
- Connectivity Optimization: A MeshCNN identifies edges for flipping to better align the mesh with the underlying geometry.
- Scalability: By operating on local patches (~2000 vertices) and fusing results globally, PRS handles meshes of arbitrary complexity.
Key Insights
PRS shifts the focus from purely geometric vertex movement to learning connectivity via edge flips. This discrete optimization is guided by a "Surface Improvement" field that targets normal consistency—the key to perceptual sharpness. Furthermore, the patch-based approach ensures resolution independence, allowing the same model to perform effectively on grids ranging from $128^3$ to $512^3$.
Experimental Results
Evaluated on the ABC Dataset (CAD models) and ShapeNet (general objects), PRS significantly outperforms existing baselines:
- Accuracy: Improved Normals F-score by 26% over Neural Dual Contouring.
- Perceptual Quality: Reduced perceptual RMSE by 42%.
- Efficiency: Produces high-quality results using 3x fewer triangles by placing geometry strategically on features.
Conclusion and Future Work
PRS demonstrates a robust method for resolution-free surface refinement. While highly effective, future work may address the "needle in a haystack" problem of identifying useful edge flips in highly refined meshes and explore end-to-end learnable solutions that eliminate remaining heuristic post-processing steps.
Summary updated