Introduction
Point clouds serve as the foundational data structure for 3D computer vision, yet real-world acquisition via LiDAR or depth sensors often results in significant noise. This noise—comprising sensor artifacts, reflections, and outliers—presents a major hurdle for downstream tasks like reconstruction and navigation. Traditional denoising models frequently rely on the assumption that noise follows a simple isotropic Gaussian distribution. However, real-world data is far more complex.
P2P-Bridge introduces a sophisticated framework that moves beyond the "noise-to-data" paradigm. By adapting Diffusion Schrödinger Bridges (DSB) to 3D point clouds, the model learns a direct, optimal transport path from noisy scans to clean surfaces, effectively bridging the gap between raw sensor output and high-fidelity geometry.
The Challenge of Real-World Noise
Most existing point cloud denoisers are trained on synthetic datasets where noise is mathematically generated. In contrast, handheld scanners produce chaotic artifacts such as "ghost points," edge flares, and non-uniform outlier clusters. Furthermore, many current methods operate purely on local geometry, ignoring the semantic context of the scene. For instance, the geometric smoothing required for a flat wall differs significantly from that of a textured rug. P2P-Bridge addresses these limitations by focusing on real-world noise distributions and incorporating high-level semantic guidance.
Methodology
The core innovation of P2P-Bridge is the application of Schrödinger Bridges to the point cloud domain, characterized by several key components:
1. Data-to-Data Diffusion
Unlike standard diffusion models that begin with pure random noise, a Schrödinger Bridge identifies the most efficient path between two known distributions: the noisy input and the clean ground truth. This "data-to-data" approach allows the model to learn the specific characteristics of sensor-specific noise rather than generic Gaussian fuzz.
2. Shortest Path Interpolation
Because point clouds are unordered sets, the diffusion process requires a tractable way to map points across time steps. The authors utilize an optimal assignment strategy—shortest path interpolation—to pair points in the noisy cloud with their counterparts in the clean cloud. This alignment ensures the model learns a meaningful geometric flow rather than a chaotic set-matching transformation.
3. PVCNN and Semantic Integration
The framework utilizes a Point-Voxel Convolutional Neural Network (PVCNN) backbone, which leverages the efficiency of voxel-based processing alongside the precision of point-based features. Crucially, the model is not limited to XYZ coordinates; it can incorporate RGB color and DINOv2 features. By projecting these high-level vision features onto the points, the model gains a semantic understanding of the scene, helping it distinguish between actual geometry and transient sensor artifacts.
Experimental Evaluation
The authors evaluated P2P-Bridge across a variety of synthetic and high-fidelity real-world datasets:
- Object-level Benchmarks: PU-Net and PC-Net were used for standard geometric evaluation.
- Scene-level Real-world Data: The model was tested on ScanNet++, which contains 330 indoor scenes with paired handheld LiDAR and high-end Faro laser scans, and ARKitScenes, featuring over 5,000 scans from Apple ARKit.
Key Insights and Results
P2P-Bridge consistently outperforms previous state-of-the-art methods such as ScoreDenoise and PD-Flow in both quantitative and qualitative metrics:
- Superior Accuracy: On the PU-Net benchmark (3% noise), P2P-Bridge achieved a Chamfer Distance (CD) of 3.99, a significant improvement over the previous best of 4.69. On the complex ScanNet++ dataset, the model (using XYZ+RGB+DINO) reached a CD of 64.44, outperforming I-PFN (70.62).
- Efficiency: By learning an optimal transport path, the model requires significantly fewer inference steps. High-quality results are achieved in just 3 to 10 steps, making it a candidate for near-real-time applications.
- Semantic Robustness: The inclusion of DINOv2 features measurably improved performance, confirming that "understanding" the scene context allows for more intelligent denoising decisions.
- Visual Quality: Qualitatively, the method produces sharper edges and smoother planar surfaces, avoiding the patchy artifacts common in other deep-learning-based denoisers.
Conclusion and Future Work
P2P-Bridge demonstrates that formulating denoising as a Schrödinger Bridge problem allows for more effective handling of real-world sensor noise. While the model excels at cleaning existing points, it is primarily a denoiser and does not currently perform point cloud completion for occluded areas. Future research may explore merging this bridge-based denoising with completion strategies to simultaneously clean and fill in missing geometry. Additionally, further optimization could push the model toward the single-step inference required for mobile augmented reality applications.
Summary updated