3D Incomplete Point Cloud Surfaces Reconstruction With Segmentation and Feature-Enhancement

Meili Wang, Ming Ju, Yuling Fan, Shihui Guo, Minghong Liao, Huijun Yang, Dongjian He, Taku KomuraIEEE Access 2019 · DOI · OpenReview

Introduction

Reconstructing 3D surfaces from raw point cloud data is a fundamental challenge in computer graphics and engineering. While 3D scanners are highly effective, the resulting datasets are often noisy, sparse, and contain significant "holes" where the scanner could not reach. Traditional reconstruction methods, such as Poisson reconstruction, frequently over-smooth these regions, leading to the loss of sharp edges and fine geometric details. Consequently, achieving high-quality models often requires intensive manual intervention. This paper proposes an automated pipeline designed to reconstruct missing parts accurately while preserving crisp geometric features with minimal human effort.

Methodology

The proposed approach utilizes a three-stage pipeline to transform raw, incomplete point clouds into high-fidelity surface meshes.

Overview of the proposed reconstruction pipeline, from input point cloud to final mesh.

1. Feature-Enhanced Pre-processing

The first stage focuses on cleaning and organizing the raw data. An anisotropic simplification operator relocates points from high-density to low-density areas, ensuring a uniform distribution without compromising the object's shape. To protect sharp boundaries, the system employs prioritized feature enhancement, selectively upsampling points at edges. Finally, normal optimization calculates point directions using an anisotropic neighborhood, preventing common errors near sharp corners.

2. Self-Adaptive Segmentation

The algorithm automatically partitions complex 3D models into simpler primitive parts. By decomposing a complex object into its constituent components (e.g., separating a hand into a palm and five fingers), the geometry becomes significantly easier to process and reconstruct.

3. Skeleton-Guided Reconstruction

For each segment, the system generates an L1-medial skeleton, which serves as a structural backbone. These skeletons guide the final surface generation via the Morfit algorithm. This structural "prior" allows the system to intelligently estimate the surface even in areas where the original scan was entirely missing data.

Results and Discussion

The researchers evaluated their method on a diverse set of models, including organic forms (Hand, Horse, Mouse), man-made objects (Vase, Laptop), and botanical structures.

Comparison of point cloud data before and after feature-enhanced pre-processing on a complex statue model.

Key Insights

  • Segmentation as a Reconstruction Prior: Segmentation is leveraged not just for classification but as a tool to maintain topological integrity. Breaking a model into parts allows the system to use skeletons to preserve the structure of individual limbs or branches.
  • Robustness to Missing Data: Medial skeletons are abstractions of a shape's center, making them inherently more robust to missing "skin" data than local neighborhood-based methods.
  • Anisotropy for Detail Preservation: By treating directions differently, the algorithm distinguishes between noise and essential geometric edges.

Performance and Efficiency

The method significantly reduces manual labor. For a complex "Mouse" model, the previous state-of-the-art (Morfit) required 12 manual user interactions to correct sweep paths, whereas the proposed method required only a single stroke. Qualitatively, the pipeline successfully captured delicate features like whiskers and eyelids that standard Poisson methods failed to reconstruct. Furthermore, the segmentation process is highly efficient, taking only 2.05 seconds for the Mouse model, outperforming deep-learning alternatives like PointNet which require extensive training.

Conclusion

This research demonstrates that combining geometric pre-processing with automated segmentation and skeletonization can effectively reconstruct incomplete 3D surfaces. While the method shows superior performance in detail preservation and automation, it can still struggle with extremely high structural complexity, where dense "webs" of parts may lead to excessive segments. Future work may involve integrating these geometric priors into end-to-end deep learning solutions to further enhance automation and handle more severe topological defects.

Summary updated

Abstract

Raw point cloud data are often noisy, superfluous, and with topological defects, such as holes. These issues cause inaccurate geometric representation in 3D reconstruction. As a result, surface reconstruction from point cloud data is a highly challenging problem. In this paper, we address the aforementioned issues by taking advantage of the embedded information of segmentation, skeletonization, and user guidance. First, we pre-process the point cloud data with three steps, relocating each point, upsampling the point data, and optimizing normals to enhance the features and geometric details; second, a segmentation method converts the input cloud into separate parts; finally, we construct curve skeletons for each part and guide the surface reconstruction with minimal user interaction, where the parts of refined smooth shapes are fused to generate the final results. The comparison studies confirmed that the proposed method is able to produce state-of-the-art results in terms of preserving sharp features, handling missing data, and requiring minimal user intervention.
BibTeX
@article{wang2019d,
  author = {Meili Wang and Ming Ju and Yuling Fan and Shihui Guo and Minghong Liao and Huijun Yang and Dongjian He and Taku Komura},
  title = {{3D Incomplete Point Cloud Surfaces Reconstruction With Segmentation and Feature-Enhancement}},
  journal = {IEEE Access},
  year = {2019},
  doi = {10.1109/access.2019.2891959},
}
Figure