Source-linked AI summary
Pointfilter: Point Cloud Filtering via Encoder-Decoder Modeling
Dongbo Zhang, Xuequan Lu, Hong Qin, Ying He
TL;DR
Point cloud filtering must remove scan noise while preserving sharp features, yet existing approaches can be fragile or tuning-intensive. Pointfilter learns local displacement vectors with an encoder-decoder network and feature-aware loss, and experiments report stronger visual and quantitative performance than state-of-the-art techniques. Its scope is limited by excessive noise and significant holes.
Problem
Point cloud filtering must remove noise while preserving sharp features, but existing methods can be less robust and require tedious parameter tuning.
Method
Pointfilter uses local noisy patches in an encoder-decoder network to learn displacement vectors toward the underlying surface, with preprocessing and a feature-aware loss.
Results
Pointfilter outperforms state-of-the-art point-set filtering techniques, or is comparable to optimization-based RIMLS, in visual quality and evaluation errors.
Takeaways & Limitations
The method automatically infers displacement vectors and produces filtered point clouds with preserved sharp features.
Takeaways & Limitations
Pointfilter has difficulty retaining sharp features under excessive noise and fails to handle significant holes in point clouds.
Abstract
from arXiv · showhide
Point cloud filtering is a fundamental problem in geometry modeling and processing. Despite of significant advancement in recent years, the existing methods still suffer from two issues: 1) they are either designed without preserving sharp features or less robust in feature preservation; and 2) they usually have many parameters and require tedious parameter tuning. In this paper, we propose a novel deep learning approach that automatically and robustly filters point clouds by removing noise and preserving their sharp features. Our point-wise learning architecture consists of an encoder and a decoder. The encoder directly takes points (a point and its neighbors) as input, and learns a latent representation vector which goes through the decoder to relate the ground-truth position with a displacement vector. The trained neural network can automatically generate a set of clean points from a noisy input. Extensive experiments show that our approach outperforms the state-of-the-art deep learning techniques in terms of both visual quality and quantitative error metrics. The source code and dataset can be found at https://github.com/dongbo-BUAA-VR/Pointfilter.
1 INTRODUCTION
Point cloud filtering is needed for noisy scan data, but existing methods can smear or poorly preserve sharp features, lack robustness, and require tedious tuning. Pointfilter addresses these issues with an end-to-end encoder-decoder approach and a feature-aware loss.
- Noisy point clouds from consumer-level depth sensors require filtering before downstream geometry, driving, and robotics applications.
- Existing techniques can smear features or lose robustness, especially under large noise or non-uniform sampling.LOP methods are not designed for sharp-feature preservation, while RIMLS and GPF depend heavily on normal filters.
- Deep-learning filters avoid tedious parameter tuning but can still smooth sharp features or generalize poorly.
- The paper compares Pointfilter with state-of-the-art methods using characteristics including normal independence and parameter requirements.
- Pointfilter uses an end-to-end encoder-decoder network and a loss function that accounts for sharp features.The network regresses displacement vectors that move noisy points toward their ground-truth positions.
2 RELATED WORK
Prior work includes optimization, projection, point-based network, and deep-learning approaches for filtering irregular point clouds. These methods use different strategies for feature preservation, local structure, and data representation, while Pointfilter uses a patch-based neural architecture.
- Point cloud filtering methods are broadly classified into two-step techniques and projection-based methods.
- Two-step based Methods: Two-step methods smooth normals and then update point positions under the filtered normals, with some variants targeting sharp features.
- Projection-based Methods: Projection-based methods include moving least squares and RIMLS, but selecting a suitable support radius can be difficult for noisy inputs.
- Point-based Network Architecture: PointNet processes raw points with shared multilayer perceptrons, providing a foundation for networks operating on irregular point-cloud data.
- Deep Learning on Point Clouds: Pointfilter's architecture uses PCA alignment, shared MLP feature extraction, max pooling, and fully connected layers to regress point displacements.
- Deep Learning on Point Cloud Filtering: Deep-learning filtering methods include height-map filtering, edge-aware losses, and projection-based smoothing, but reported issues include manual labeling, limited local information, and weak feature preservation.
3 METHOD
Pointfilter filters noisy point clouds locally by learning displacement vectors from canonicalized neighborhoods. Its encoder-decoder architecture uses projection- and repulsion-aware losses to approximate the underlying surface, preserve sharp features, and maintain point distribution.
- Overview: Pointfilter formulates filtering as learning a displacement vector that projects each noisy point onto the underlying surface.The network treats additive noise as the displacement between noisy and clean point clouds.
- Preprocessing: Each local patch is translated, scaled, and PCA-aligned to reduce unnecessary degrees of freedom and improve invariance to rigid transformations.The patch is centered at the noisy point, normalized by its radius, and aligned to canonical axes using PCA.
- Network Architecture: The encoder extracts multi-scale neighborhood features and aggregates them into a latent representation, while the decoder regresses the displacement vector.The architecture uses shared MLP feature extractors, max pooling, and fully connected regression layers.
- Loss Function: The projection loss incorporates neighboring normals through a bilateral mechanism because closest-point and basic projection losses can blur sharp features.Nearby points receive Gaussian distance weights, while normal similarity helps preserve feature structure.
- Loss Function: A repulsion term penalizes point aggregation, and the authors set its trade-off parameter η to 0.97 during training.The term is intended to keep filtered points relatively uniformly distributed.
- Network Inference: During inference, each aligned noisy patch produces a displacement vector that is mapped back to the original space, with multiple iterations used for larger noise.The filtered point and noisy point are related through the transformed network output and inverse PCA rotation.
4 EXPERIMENTAL RESULTS
Experiments evaluate Pointfilter on synthetic and raw scanned point clouds against classical and learning-based methods, using visual comparisons, error metrics, and robustness tests. The method generally preserves sharp features, removes noise effectively, and achieves the lowest average errors on the synthetic test set.
- 4.1 Dataset: The training dataset contains 132 models from 22 clean CAD and non-CAD models, while testing uses 15 synthesized models and 7 raw scanned point clouds.Training noise levels range from 0.0% to 2.5%; synthesized test noise levels range from 0.5% to 2.5%.
- 4.2 Comparisons and 4.3 Evaluation Metrics: Pointfilter is compared with WLOP, CLOP, RIMLS, GPF, EC-Net, PCN, and TotalDenoising using distance error, Chamfer distance, MSE, and P2F metrics.The evaluation defines distance errors from nearest filtered neighbors and also uses Chamfer distance and point-to-surface distance.
- 4.4 Visual Comparisons: At 0.5% synthetic noise, Pointfilter produces visually better noise removal and feature preservation than the compared methods, while RIMLS and GPF become less robust at larger noise.The comparison covers four CAD models and reports gaps and outliers in RIMLS and GPF results at 1.0% noise.
- 4.4 Visual Comparisons: Pointfilter generalizes to impulsive, uniform, and raw scanned noise without retraining for most raw-noise experiments, retaining sharp features in visual comparisons.The reported raw-noise tests include scanned shapes, scenes, and virtually scanned models.
- 4.5 Quantitative Comparisons: Across 15 synthetic test models, Pointfilter achieves the lowest average CD, MSE, and P2F errors and is reported as generally the most accurate and easiest to use.RIMLS achieves comparable results but requires quality normals and trial-and-error parameter tuning.
- 4.6 Robustness: Pointfilter remains especially robust at 2.5% noise and generally outperforms competitors across point densities, except for the 10K-point case.The authors attribute the sparse-point limitation to insufficient neighboring information for depicting local structure.
5 CONCLUSION
Pointfilter is an automatic framework for feature-preserving point cloud filtering that performs competitively across visual and error-based evaluations. Its main limitations are excessive noise, significant holes, and limited suitability for large-scale outlier removal.
- Pointfilter automatically infers displacement vectors to filter noisy point clouds while preserving sharp features.The framework is trained end-to-end and generates filtered points from noisy inputs.
- Extensive experiments show visual and evaluation-error performance that outperforms state-of-the-art point set filtering methods or matches RIMLS under its tuning requirements.RIMLS requires quality normals and trial-and-error parameter tuning, whereas Pointfilter is automatic.
- Pointfilter is complementary to PCN because it targets sharp-feature preservation rather than large-scale outlier removal.The paper explicitly distinguishes Pointfilter from PCN's emphasis on heavy outliers removal.
- Pointfilter struggles to retain sharp features under excessive noise and fails to handle significant holes in point clouds.The authors propose incorporating global shape information in future work.