CurveCloudNet: Processing Point Clouds with 1D Structure

TL;DR

CurveCloudNet tackles semantic segmentation and object understanding in 3D point clouds by exploiting the 1D curve structure inherent in laser scans. It introduces a curve cloud representation—polylines derived from per-beam lidar sweeps—and a set of 1D curve operations (1D FPS, curve grouping, symmetric curve convolution) that are interleaved with traditional point-based processing in a U-Net–style backbone. The approach demonstrates strong, scalable performance across ShapeNet, KortX, A2D2, nuScenes, and KITTI, outperforming or matching state-of-the-art point-based and voxel-based methods while better handling large outdoor scenes and diverse scanning patterns. Together, the curve-centric paradigm provides a flexible, efficient backbone for 3D perception in open-world robotics, with clear avenues for extending curve reasoning to broader data types and future integrations such as curve-to-curve attention and cross-curve interactions.

Abstract

Modern depth sensors such as LiDAR operate by sweeping laser-beams across the scene, resulting in a point cloud with notable 1D curve-like structures. In this work, we introduce a new point cloud processing scheme and backbone, called CurveCloudNet, which takes advantage of the curve-like structure inherent to these sensors. While existing backbones discard the rich 1D traversal patterns and rely on generic 3D operations, CurveCloudNet parameterizes the point cloud as a collection of polylines (dubbed a "curve cloud"), establishing a local surface-aware ordering on the points. By reasoning along curves, CurveCloudNet captures lightweight curve-aware priors to efficiently and accurately reason in several diverse 3D environments. We evaluate CurveCloudNet on multiple synthetic and real datasets that exhibit distinct 3D size and structure. We demonstrate that CurveCloudNet outperforms both point-based and sparse-voxel backbones in various segmentation settings, notably scaling to large scenes better than point-based alternatives while exhibiting improved single-object performance over sparse-voxel alternatives. In all, CurveCloudNet is an efficient and accurate backbone that can handle a larger variety of 3D environments than past works.

Figures (10)

• Figure 1: Visualizations of the input curve cloud (left) and CurveCloudNet's segmentation prediction (right) for each of our five evaluation datasets. Each evaluation dataset exhibits distinct size, structure, and laser scanning pattern, as shown in \ref{['tab:summary-results-and-datasets']}
• Figure 2: Overview of Curve Cloud Reasoning. Starting from laser-scanned input data, we ① link points into polylines to ② parameterize the point cloud as a curve cloud (see \ref{['sec:prelims']}). We develop operations for learned architectures to specifically exploit the curve structure, including ③ 1D farthest-point-sampling along a curve, ④ curve grouping, and ⑤ symmetric curve convolutions (see \ref{['sec:curve-ops']}).
• Figure 3: CurveCloudNet Architecture. The network employs a mix of curve and point layers to process a curve cloud through progressive down-sampling followed by up-sampling with skip connections. Curve layers operate on higher resolutions to efficiently capture the 1D structure, while at lower resolutions point layers propagate information across curves. Feature dimensions are listed above each block.
• Figure 4: Qualitative Results on Kortx. CurveCloudNet successfully segments fine-grained parts by leveraging curve structures.
• Figure S1: Point to Curve Cloud Conversion. Algorithm (in Python) to convert an input point cloud into a set of polylines.