Random screening-based feature aggregation for point cloud denoising

Feature aggregation is a core operation in learning-based point cloud denoising: for each noisy point, features from its neighbourhood are pooled to estimate a clean position. The quality of aggregation depends critically on which neighbouring points are included—outlier points and locally dense clusters can dominate the pooled feature and pull the estimate off the true surface. This paper proposes random screening as a lightweight solution: instead of aggregating over the full neighbourhood, multiple random subsets are drawn and screened by a consistency criterion that favours subsets whose geometric statistics match the expected local surface model. Only the highest-scoring subsets contribute to the final aggregated feature. This stochastic screening process is differentiable and integrable into existing denoising architectures as a drop-in replacement for standard neighbourhood aggregation. No additional training data or network redesign is required. Evaluated on ShapeNet3D and real LiDAR scan benchmarks and published in Computers & Graphics (2023), random screening-based aggregation consistently improves Chamfer distance and point-to-mesh accuracy over deterministic aggregation, with the largest gains on point clouds with high outlier rates.

Problem setting

Point cloud denoising methods based on feature aggregation compute noise-free position estimates by aggregating information from local neighbourhoods. Fixed neighbourhood selection is, however, sensitive to outlier points and non-uniform sampling, which can corrupt the aggregated features. This paper introduces random screening-based feature aggregation, where multiple random subsets of each neighbourhood are sampled and screened before aggregation.

In the broader publication record, this work appears in Computers & Graphics, 116:64–72. The visual notes below pair the paper’s original figures with a concise reading of the method, experimental setup, and reported results.

Method and visual evidence

The method works on 3D geometric observations such as point clouds, poses, correspondences, or segmented regions, then uses the proposed representation to improve robustness under noise, viewpoint change, or limited observations.

The extracted figures below show the geometric representation, network or optimization pipeline, and qualitative or quantitative results.

Method overview. This image is extracted from an embedded PDF image object on page 3, then recomposed for web display.

Representation and setup. This image is extracted from an embedded PDF image object on page 5, then recomposed for web display.

Experimental evidence. This image is extracted from an embedded PDF image object on page 5, then recomposed for web display.

Result comparison. This image is extracted from an embedded PDF image object on page 6, then recomposed for web display.

Additional visual result. This image is extracted from an embedded PDF image object on page 7, then recomposed for web display.

Results and impact

The evaluation reported in Computers & Graphics, 116:64–72 uses the extracted figures above to show the method’s measurement, reconstruction, segmentation, matching, or diagnostic behavior on representative experiments. These visuals are paired with the paper’s quantitative or qualitative analysis to make the workflow easier to inspect from the homepage.

Source handling

I extracted 8 candidate image objects from paper.pdf and generated the compressed WebP figures used on this page.