Joint Bilateral Mesh Denoising using Color Information and Local Anti-Shrinking

Oliver Wasenmüller, Gabriele Bleser, Didier StrickerJ. WSCG 2015 · OpenReview

Abstract

3D mesh denoising is a critical step in processing data from depth cameras and photogrammetry, yet traditional methods often suffer from volume shrinkage and the erosion of sharp geometric features. This paper introduces Joint Bilateral Mesh Denoising (JBMD), a framework that leverages color information as a proxy for geometric boundaries and implements a local anti-shrinking mechanism. By integrating color similarity into the bilateral weighting process and applying targeted vertex corrections, JBMD effectively removes noise while preserving fine details and maintaining the original volume of the model. Experimental results demonstrate that JBMD outperforms standard bilateral and Laplacian filters in both quantitative error metrics and qualitative visual fidelity.

Introduction

As 3D scanning technology becomes more accessible, the demand for high-quality mesh reconstruction grows. However, raw scans are frequently contaminated with noise—artificial irregularities that obscure the true surface of the object. While denoising filters are designed to smooth these surfaces, they face two primary challenges: the "shrinking problem," where iterative smoothing causes convex regions to migrate inward, and the neglect of auxiliary data, such as color textures, which can provide vital clues about the underlying geometry.

This paper addresses these limitations by proposing a "color-aware" filtering approach combined with a surgical, local correction step to prevent volume loss.

Methodology

Joint Bilateral Filtering with Color Guidance

The proposed Joint Bilateral Mesh Denoising (JBMD) algorithm extends the standard bilateral filter by incorporating color data. In a traditional bilateral filter, a vertex's new position is calculated based on its neighbors, weighted by physical distance and distance from the vertex's tangent plane. JBMD introduces a third weight: the difference in color.

This approach is based on the principle that sharp color transitions often align with sharp geometric edges. By reducing the influence of neighbors with significantly different colors, the filter prevents the smoothing out of distinct geometric features that might otherwise be blurred by noise-reduction processes.

Local Anti-Shrinking

Smoothing filters naturally pull vertices toward the center of a local neighborhood, leading to a "melting" effect in convex areas like the tip of a nose or the edge of a mechanical part. While traditional methods attempt to fix this via global volume preservation (scaling the entire model back up), JBMD utilizes a local anti-shrinking method. It calculates a "weighted mean signed shift" to identify exactly where the mesh has migrated inward and applies a counter-shift only to those specific vertices. This ensures that flat areas remain smooth while high-curvature features retain their original volume.

Iterative Refinement

The algorithm operates iteratively, automatically adjusting its sensitivity over time. By tightening the $\sigma_o$ parameter as the process progresses, the algorithm moves from broad-brush noise removal to fine-detail refinement, ensuring a stable convergence that avoids over-smoothing the model into a featureless sphere.

Experimental Evaluation

Datasets

The authors validated JBMD across three categories of 3D data:

  1. Synthetic Meshes: Using the fandisk model with artificial noise to establish a ground truth for error measurement.
  2. High-Precision Scans: Complex organic shapes, including lion and allegorie statues, were used to test the algorithm's performance on intricate surfaces.
  3. Real-World Reconstructions: Scans of human heads produced via Agisoft PhotoScan, representing the unpredictable noise patterns found in consumer-grade hardware.

Results and Discussion

Quantitative analysis on the fandisk model showed that JBMD achieved a Root Mean Square Error (RMSE) of approximately 0.048, a marked improvement over the ~0.066 achieved by standard Bilateral Mesh Denoising (BMD).

Qualitatively, JBMD demonstrated superior feature preservation in real-world scans. In human face models, JBMD successfully maintained the distinct shapes of eyebrows, nostrils, and lips, whereas competing filters tended to flatten these features. Error maps confirmed that JBMD keeps high-error zones confined to actual sharp edges, whereas other methods distribute error across broader regions of the mesh.

Conclusion

Joint Bilateral Mesh Denoising offers a robust solution for 3D reconstruction pipelines by treating color as a first-class citizen in the denoising process. The combination of color-guided weighting and local anti-shrinking allows for high-fidelity results that preserve both the volume and the sharp details of the original object. Future work may focus on optimizing the computational cost of color-weight calculations, potentially through GPU acceleration, to enable real-time applications in live 3D scanning environments.

Summary updated

BibTeX
@article{wasenmuller2015joint,
  author = {Oliver Wasenmüller and Gabriele Bleser and Didier Stricker},
  title = {{Joint Bilateral Mesh Denoising using Color Information and Local Anti-Shrinking}},
  journal = {J. WSCG},
  year = {2015},
  url = {https://openreview.net/forum?id=NQ2_gbR1Ga},
}
Figure