TL;DR
This paper introduces a novel method for completing damaged 3D models by leveraging mirror symmetry. Unlike deep learning approaches that struggle with unseen object types, this method uses local geometric features like curvatures and edges to robustly identify symmetry planes and reconstruct missing parts, achieving competitive results.
What Problem Does This Paper Address?
Modern 3D scanning technologies, while impressive, often produce incomplete or flawed 3D models due to issues like occlusions, poor lighting, or scanner imperfections. Point cloud completion is a crucial post-processing step to fix these defects. While deep learning methods have shown promise, they often require extensive training data for specific object categories and perform poorly on novel, unseen objects. This paper addresses the need for a more generalizable and robust method for 3D object completion, particularly for objects exhibiting mirror symmetry, which is common in man-made environments.
Methods
The proposed method employs a four-stage pipeline:
- 3D Model Representation: The input is a 3D mesh, which is converted into a point cloud. Local geometric features, specifically principal curvatures and directions, are computed for each point.
- Symmetry Plane Detection (Curvature-Based): This approach matches points with similar principal curvatures. Pairs of points with compatible curvatures and principal directions are used to hypothesize symmetry planes. These planes are then validated by checking if mirrored local patches align, using the Mean Shift algorithm for clustering.
- Symmetry Plane Detection (Edge-Based): This method is designed for objects with planar features where curvature might be less informative. It detects edge points in the point cloud and fits lines to these points. Pairs of parallel or intersecting lines are used to define potential symmetry planes.
- Symmetry Plane Voting and Completion: Both curvature-based and edge-based methods generate candidate symmetry planes with associated weights. A voting mechanism, often involving Mean Shift clustering on these planes, is used to determine the single "best" symmetry plane. This plane is then used to mirror the existing parts of the object to fill in the damaged or missing regions.
The novelty lies in the combination of curvature and edge-based local geometric feature matching for symmetry detection, and the robust voting mechanism to select the most likely symmetry plane.
Datasets
The primary dataset used for evaluation is ModelNet40, a widely used benchmark for 3D object recognition. The authors selected 5 visually symmetric objects from each of the 40 categories, resulting in a dataset of 200 objects. For evaluation, meshes were converted into point clouds, and artificial damage was introduced by removing approximately 15% of the surface area.
Key Insights and Ideas
- Local Geometry is Key: The paper strongly advocates for using local geometric features (curvatures, edges) as a robust basis for symmetry detection, especially for objects not seen during training.
- Complementary Symmetry Detection: Combining curvature-based and edge-based methods provides a more comprehensive approach, as each excels in different scenarios (smooth surfaces vs. sharp edges/planes).
- Robustness through Voting: The Mean Shift clustering and weighted voting for symmetry planes effectively filters out spurious candidates and identifies the most dominant symmetry.
- Generalizability: By relying on geometric properties rather than learned features from specific classes, the method aims for better performance on unseen 3D objects.
Key Results
The paper reports competitive results when comparing their method against traditional symmetry-based methods and deep learning approaches (MSN, PoinTr, GRNet) on the ModelNet40 dataset. The "Improvement Rate" metric, which quantifies how much better the completed object is compared to the damaged one, shows that their combined approach (Our Combined) achieved a mean improvement rate of 82.1% (with a median of 26.4%). Their curvature-based approach alone also performed well, with a mean improvement rate of 71.8%. The authors note that their method performed similarly to the traditional method from [1], and generally better than some deep learning methods when tested on unseen data.
Limitations and Future Work
The authors acknowledge several limitations:
- Mirror Symmetry Requirement: The method is strictly limited to objects possessing mirror symmetry.
- Symmetry of Missing Parts: If missing parts are themselves mirror symmetric or have a large intersection with their mirrored counterparts, reconstruction can be problematic.
- Parameter Sensitivity: The performance can be significantly affected by the choice of parameters, requiring fine-tuning for individual objects, which impacts robustness.
- Poor Symmetry Detection: If the symmetry plane is not found accurately, the completion quality suffers.
Future work could explore extending the method to handle other types of symmetries or developing more automated parameter tuning strategies to improve robustness.
Summary updated