Abstract
Raw point clouds captured by LiDAR or depth cameras are inherently noisy, necessitating denoising for downstream tasks like autonomous driving and 3D modeling. Traditional deep-learning denoising methods often suffer from "attention bias," where the network aggressively processes already-clean points to minimize loss, leading to over-smoothing and the loss of fine geometric details. We present the Adaptive Stop Denoising Network (ASDN), which introduces an "early exit" strategy. By utilizing an entropy-based "recoverability factor," ASDN identifies points that are sufficiently restored and stops their processing, preserving sharp edges and intricate structures. Our results demonstrate that ASDN consistently outperforms state-of-the-art methods across synthetic and real-world benchmarks.
Introduction: The Problem of Over-Denoising
Existing deep-learning denoising frameworks typically treat all points in a cloud equally throughout the entire inference process. However, high-noise points offer a greater potential for loss reduction during training, causing neural networks to focus disproportionately on them. This "greedy" behavior leads to a fundamental flaw: the network continues to apply aggressive denoising cycles to points that have already reached a clean state. The result is over-denoising, where sharp corners are rounded and thin structures, such as chair legs or fine textures, are wiped away. The challenge lies in creating a mechanism that knows exactly when to stop.
Methodology
Point Cloud Entropy and Recoverability
To quantify the quality of a point cloud patch, the authors adapt Shannon entropy to 3D space. By voxelizing the point cloud and calculating distribution density, they derive a metric for "messiness." From this, they define the Recoverability Factor ($\rho$), which is the ratio of entropy between the noisy point cloud and its clean counterpart. A lower $ ho$ indicates a more challenging restoration task, providing a robust quality control metric that is invariant to point density and resolution.
The Adaptive Stop Denoising Network (ASDN)
ASDN is built on a U-Net style architecture integrated with a specialized Adaptive Classifier (AC). The AC acts as the system's "brain," using EdgeConv layers to predict the recoverability factor of a point patch. Based on this prediction, the classifier maps the patch to a specific layer ($\lambda$) in the U-Net.
If the classifier determines a patch is "clean enough," it triggers an early transition from the encoder to the decoder, bypassing further aggressive denoising cycles. This dynamic stopping ensures that the underlying manifold of the surface is preserved rather than melted by excessive processing.
Architecture Components
The network's effectiveness is further bolstered by two key modules:
- Local Feature Aggregation (LA): Used in the encoder to capture neighborhood geometry and local structures.
- Multi-Head Cross-Attention (MHCA): Employed in the decoder to correlate features between the encoding and decoding stages, ensuring global consistency during reconstruction.
Experimental Results
Quantitative Performance
ASDN was evaluated against state-of-the-art (SOTA) methods including IterativePFN, PDFlow, and Score. On the PUNet 10K benchmark with 3% Gaussian noise, ASDN achieved a Chamfer Distance (CD) of 3.079, significantly outperforming the next best competitor, DeepPSR (4.075). Ablation studies confirmed that removing the Adaptive Classifier (the "w/o AC" test) led to a sharp performance drop, validating the early-stopping strategy.
Real-World Robustness
The model's performance extends beyond synthetic data. When tested on real-world scans from the Kinect v2 and the large-scale Paris-rue-Madame street scene dataset, ASDN achieved the lowest CD (18.22) and demonstrated superior ability to handle sensor-specific noise and "in-the-wild" scenarios.
Qualitative Analysis
Visual comparisons highlight ASDN's ability to preserve thin structures and sharp corners that other methods typically over-smooth. By intentionally stopping the denoising process for specific points, the model maintains the fidelity of the original geometric intent.
Conclusion and Future Work
ASDN demonstrates that in the context of point cloud denoising, "less is more." By selectively processing points based on their recoverability, the model avoids the pitfalls of over-smoothing inherent in fixed-iteration networks. Future research will explore the theoretical mathematical relationship between entropy and surface recoverability, as well as the application of adaptive stopping to other 3D tasks such as point cloud completion and normal estimation.
Summary updated