Point cloud super-resolution improves 3D LiDAR imaging quality outdoors
Local-Global Joint Modeling for Point Cloud Super-Resolution Guided by Multi-Scale Spatial Priors
Computer Vision and Pattern Recognition
Summary
Sparse 3D point clouds from outdoor LiDAR scans often miss detail, especially far away. The authors propose a method that combines local and global information with spatial cues at multiple scales to better enhance these point clouds. Their technique also reduces noise and sharp artifacts that happen when converting 2D images back to 3D. Tests on real-world datasets show it reconstructs 3D scenes with better accuracy.
What this means in practice
- •For autonomous vehicle teams: Enhance sparse outdoor LiDAR scans to improve environment perception accuracy for autonomous driving.
- •For robotics developers: Improve 3D mapping quality in outdoor robotics by refining range-image-based point cloud data.
Authors
Wang Yang, Wenyi Ge
Abstract
Range-image-based point cloud super-resolution provides an effective way to reconstruct sparse outdoor point clouds by projecting low-resolution LiDAR scans into a regular 2D representation. However, current methods still face three challenges. First, range images exhibit pronounced vertical nonuniformity, with sparse observations in distant regions and denser observations in nearby regions. Some cross-scale fusion mechanisms do not account for the varying reliability of features at different vertical positions, allowing invalid values and noise from distant sparse regions to propagate into the decoder. Second, a single convolutional or sequence-modeling architecture often struggles to preserve both local high-frequency geometric details and global long-range dependencies, resulting in over-smoothed reconstructions. Third, smooth transition depths near foreground-background boundaries can produce spike artifacts when the 2D range image is back-projected into 3D. To address these issues, we propose a point cloud super-resolution method based on multi-scale spatial priors and joint local-global modeling. The design comprises a spatially gated fusion module, a multi-granularity local-global cooperative perception module, and 2D median-residual denoising. Experiments on KITTI-360 and nuScenes demonstrate effective reconstruction, with IoU scores of 0.4703 and 0.3593, respectively, supporting the feasibility of the proposed method for range-image super-resolution.