Point Cloud Denoising Based on Tensor Tucker Decomposition

Jianze Li, Xiao-Ping Steven Zhang, Tuan TranICIP 2019 · DOI · OpenReview

Abstract

3D point cloud denoising is a fundamental task in computer vision and remote sensing, essential for applications such as autonomous navigation and digital heritage preservation. Raw point clouds are frequently corrupted by noise, particularly in challenging scanning environments. This paper presents the TUDE (Tucker Decomposition-based point cloud Denoising) algorithm, a non-local approach that exploits the self-similarity of local surface patches. By organizing similar patches into a 3rd-order tensor and applying Tucker decomposition, the method filters noise in a transformed "frequency" domain. Experimental evaluations demonstrate that TUDE is highly effective in high-noise scenarios, offering competitive performance and superior visual fidelity compared to state-of-the-art methods.

Introduction

The proliferation of 3D scanning technologies like LiDAR and depth cameras has made point clouds a primary data format for representing 3D geometry. However, sensor noise and environmental factors often degrade the quality of these scans. Traditional denoising methods, such as Moving Least Squares (MLS), focus on local surface smoothing but often fail when noise levels are high, leading to blurred features or artifacts. Recent graph-based methods offer improvements but can be computationally intensive or sensitive to initial conditions.

The authors address these challenges by proposing a non-local denoising framework. By identifying and grouping similar patches across the entire point cloud, the algorithm leverages redundant structural information to distinguish between true geometric features and random noise.

The TUDE Algorithm

The TUDE algorithm follows a multi-stage pipeline designed to maximize the benefits of higher-order data representations:

  1. Local Patch Extraction: For each seed point in the noisy cloud, a local surface patch is formed by its $K$ nearest neighbors.
  2. Similarity Grouping via ICP: To exploit non-local redundancy, the algorithm searches for patches with similar geometric structures. The Iterative Closest Point (ICP) cost function is used to measure similarity, allowing for the grouping of patches that may have different orientations.
  3. Tensor Construction: The grouped patches are stacked to form a 3rd-order tensor. This representation captures both the local geometry of individual patches and the shared structure across the group.
  4. Tucker Decomposition: The algorithm applies Tucker decomposition (using the HOOI method) to the patch tensor. This decomposes the data into a core tensor and factor matrices, effectively projecting the patches into a custom coordinate system.
  5. Hard Thresholding: In this transformed domain, the core tensor represents the "frequencies" of the geometric signal. Small coefficients, which typically correspond to noise, are zeroed out through hard thresholding.
  6. Reconstruction and Aggregation: The denoised patches are transformed back to their original space and averaged to produce the final clean point cloud.

Experimental Results

The performance of TUDE was evaluated on three standard models: Gargoyle, DC (architectural), and Daratech (mechanical). These models were subjected to varying levels of Gaussian noise to simulate real-world scanning conditions.

Quantitative Performance

Quantitative analysis using Mean Square Error (MSE) shows that TUDE is particularly robust as noise levels increase. While local methods and some graph-based approaches like GLR perform well in low-noise settings ($\sigma = 0.04$), TUDE demonstrates superior stability in high-noise environments ($\sigma = 0.1$). For instance, on the Gargoyle model at $\sigma = 0.1$, TUDE reached an MSE of 0.416, surpassing the 0.436 recorded by the GLR algorithm.

Table 1: Comparative MSE results for various models and denoising methods across different noise levels.

Qualitative and Visual Quality

Visually, TUDE produces denoised models that are more compact and preserve sharp features better than competing methods. The non-local grouping ensures that even under heavy noise, the underlying structural patterns are recovered accurately.

Figure 1: Visual comparison of point cloud denoising: (a) Ground truth model, (b) Noisy input, and (c) Denoised result using the TUDE algorithm.

Discussion and Conclusion

The TUDE algorithm represents a significant step forward in non-local point cloud denoising. By treating groups of similar patches as tensors, it effectively utilizes the inherent self-similarity of 3D objects. The use of Tucker decomposition provides a powerful mechanism for noise separation that is more robust than simple rank-1 approximations.

While TUDE is highly effective for high-noise data, it remains sensitive to parameter selection, such as patch size and search range. Future work will focus on automating parameter tuning and exploring nonlinear tensor models to further enhance denoising performance in complex geometric regions.

Summary updated

Abstract

In this paper, we propose a new algorithm for point cloud de-noising based on the tensor Tucker decomposition. We first represent the local surface patches of a noisy point cloud to be matrices by their distances to a reference point, and stack the similar patch matrices to be a 3rd order tensor. Then we use the Tucker decomposition to compress this patch tensor to be a core tensor of smaller size. We consider this core tensor as the frequency domain and remove the noise by manipulating the hard thresholding. Finally, all the fibers of the denoised patch tensor are placed back, and the average is taken if there are more than one estimators overlapped. The experimental evaluation shows that the proposed algorithm outperforms the state-of-the-art graph Laplacian regularized (GLR) algorithm when the Gaussian noise is high (σ = 0.1), and the GLR algorithm is better in lower noise cases (σ = 0.04, 0.05, 0.08).
BibTeX
@inproceedings{li2019point,
  author = {Jianze Li and Xiao-Ping Steven Zhang and Tuan Tran},
  title = {{Point Cloud Denoising Based on Tensor Tucker Decomposition}},
  booktitle = {ICIP},
  year = {2019},
  doi = {10.1109/icip.2019.8803602},
}
Figure