TL;DR
This paper introduces a novel method for reconstructing complete 3D models from incomplete point clouds by leveraging detected primitive shapes (like planes and cylinders) to guide the filling of holes. Their approach uses a graph-cut optimization framework to synthesize plausible geometry, including sharp features from primitive intersections, and can also produce idealized, CAD-friendly models.
What Problem Does This Paper Address?
Acquiring 3D data using scanners or stereo vision often results in incomplete models due to occlusions, sensor limitations, or material properties. These "holes" in the surface representation hinder further processing, rendering, and analysis. While existing methods attempt to fill these gaps using general smoothness assumptions or example-based approaches, they often struggle to reconstruct sharp features or accurately infer the underlying geometric structure, especially when multiple primitives intersect. This paper tackles the challenge of robustly completing these incomplete models by explicitly using detected geometric primitives as a strong prior for the reconstruction process.
Methods
The core of the proposed method is a surface energy functional that guides the reconstruction. This functional incorporates:
- Primitive Adherence: Encourages the reconstructed surface to align with detected primitive shapes. This is achieved by defining a vector field derived from primitive normals and penalizing deviations.
- Constraint Enforcement: Ensures the reconstructed surface respects known inside/outside regions of the original object.
- Primitive Connectivity: A crucial novel aspect, this constraint ensures that all surface parts corresponding to a single primitive form a connected component. This prevents undesirable "shortcuts" across holes.
The continuous energy functional is discretized and then minimized using an efficient graph-cut algorithm. For the primitive connectivity constraint, which is not directly amenable to graph-cuts, they introduce a greedy optimization strategy. This iterative process refines the reconstruction by penalizing disconnected primitive regions until convergence.
Additionally, the paper proposes a method for detail reconstruction by combining their primitive-guided approach with a detail-preserving functional (similar to [LB07]). This allows for the recovery of fine surface features that might not be well-represented by primitives, while still using primitives to fill larger holes. Finally, an extended marching cubes algorithm is used for surface extraction, leveraging the primitive information to accurately recover sharp features at primitive intersections.
Datasets
The paper evaluates its method on several synthetic and real-world datasets:
- Fandisk model: A synthetic example used to demonstrate the effects of connectivity enforcement and consistent edge labeling.
- Height field examples: Used to showcase the method's ability to handle complex intersections and compare against existing range image completion methods.
- Carter model: A real-world example reconstructed from nine registered range scans, demonstrating the handling of numerous holes and primitives.
- Master cylinder model: Used to illustrate detail reconstruction, where parts are not well-approximated by primitives.
- Oil-pump model: A complex reconstruction from nine scans, highlighting the method's ability to handle a large number of primitives and significant missing geometry.
No new benchmarks were introduced; rather, existing models and datasets were used for evaluation.
Key Insights and Ideas
- Primitives as Strong Priors: The paper demonstrates that explicitly using detected geometric primitives (planes, cylinders, etc.) as a prior significantly improves the robustness and plausibility of 3D model completion, especially in the presence of complex intersections and large holes.
- Connectivity is Key: The introduction of a primitive connectivity constraint is a critical insight. It ensures that the completed surface parts belonging to a single primitive are logically connected, preventing geometrically implausible shortcuts.
- Graph-Cut for Geometry Inference: The successful application of graph-cut optimization to infer missing geometry guided by primitives is a powerful demonstration of its versatility beyond simple segmentation or labeling tasks.
- Hybrid Reconstruction: The ability to seamlessly combine primitive-guided completion with detail-preserving reconstruction offers a flexible approach for handling diverse object types and scan qualities.
- Idealized Models for CAD: The method's capability to produce idealized, primitive-based reconstructions makes the output directly usable in CAD systems, bridging the gap between scanned data and design workflows.
Key Results
- The proposed method successfully reconstructs complete and plausible 3D models from incomplete point clouds, even in challenging cases with multiple intersecting primitives and large holes, outperforming existing methods in these scenarios.
- The primitive connectivity constraint effectively prevents undesirable reconstructions where holes are "cut off" by disconnected primitive parts.
- The consistent edge labeling mechanism ensures smooth transitions between primitives, avoiding "bumpy" appearances at intersections.
- The detail reconstruction variant successfully preserves fine surface features while still leveraging primitives for hole filling.
- The method can generate idealized, primitive-based models suitable for CAD applications.
- Experimental results show convergence in less than fifteen iterations for the greedy connectivity enforcement, and the method scales to models with hundreds of primitives.
Limitations and Future Work
- Symmetry Handling: The current method does not explicitly incorporate symmetry detection, meaning it cannot infer missing geometry on one side of an object based on the structure present on the opposite side.
- Primitive Detection Dependency: The quality of the reconstruction is dependent on the accuracy and completeness of the initial primitive detection.
- Tiny Artifacts: Misaligned primitives can lead to small gaps that are filled with minimal surfaces, potentially introducing minor artifacts. The authors suggest a refitting algorithm could alleviate this.
- Limited Primitive Set: While the current set of primitives (planes, spheres, cylinders, cones, tori) is sufficient for many cases, extending the method to more complex primitives like NURBS is suggested for future work.
- Combining with Other Priors: The authors envision combining their primitive-based approach with methods based on self-similarity (like symmetry detection) or texture synthesis to handle an even wider range of challenging completion scenarios.
Summary updated