SRMamba: Mamba for Super-Resolution of LiDAR Point Clouds

TL;DR

SRMamba tackles the problem of upsampling sparse LiDAR point clouds from novel viewpoints by operating in the range-view while preserving 3D structure. It combines Hough Voting and Hole Compensation to improve the quality of range images, with a Visual State Space Model–based bidirectional scanning encoder–decoder and an asymmetric U‑Net to capture long-range dependencies and adapt to multi-beam LiDARs. Key contributions include the Hough Voting + Hole Compensation module, the VSSM‑driven SS2D attention-free global dependency mechanism, and an asymmetric architecture that handles varying beam counts, validated on SemanticKITTI and nuScenes with strong quantitative and qualitative gains. The approach offers improved geometric fidelity and robustness for large-scale, low-density reconstructions at reduced computational cost, enabling practical deployment in autonomous systems and scene understanding.

Abstract

In recent years, range-view-based LiDAR point cloud super-resolution techniques attract significant attention as a low-cost method for generating higher-resolution point cloud data. However, due to the sparsity and irregular structure of LiDAR point clouds, the point cloud super-resolution problem remains a challenging topic, especially for point cloud upsampling under novel views. In this paper, we propose SRMamba, a novel method for super-resolution of LiDAR point clouds in sparse scenes, addressing the key challenge of recovering the 3D spatial structure of point clouds from novel views. Specifically, we implement projection technique based on Hough Voting and Hole Compensation strategy to eliminate horizontally linear holes in range image. To improve the establishment of long-distance dependencies and to focus on potential geometric features in vertical 3D space, we employ Visual State Space model and Multi-Directional Scanning mechanism to mitigate the loss of 3D spatial structural information due to the range image. Additionally, an asymmetric U-Net network adapts to the input characteristics of LiDARs with different beam counts, enabling super-resolution reconstruction for multi-beam point clouds. We conduct a series of experiments on multiple challenging public LiDAR datasets (SemanticKITTI and nuScenes), and SRMamba demonstrates significant superiority over other algorithms in both qualitative and quantitative evaluations.

Figures (13)

• Figure 1: Super resolution diagram of the point cloud. The top shows the original 16-line sparse point cloud, with low point density and blurry object outlines; the bottom shows the 64-line point cloud after super-resolution processing, with significantly higher point density, and the structure and details of the object can be clearly reproduced, more accurately reflecting the 3D geometry of the real scene.
• Figure 2: Limitations of point cloud super-resolution based on traditional range-view. (1) left: horizontal linear hole. (2) right: offset in the new view.
• Figure 3: Overall framework. The present method takes a sparse point cloud as input, generates a range image, employs a U-Net structure for feature extraction and generates a high-resolution image, back-projects it into the 3D space, and finally generates a high-resolution, high-fidelity representation of the point cloud.
• Figure 4: SRMamba adopts a hierarchical encoder-decoder architecture, with VSS blocks, downsampling, and PixelShuffle as its core building components. By leveraging multi-scale feature fusion, the model performs super-resolution upsampling tailored to range images.
• Figure 5: A bidirectional scanning mechanism in the spatial domain with scanning directions including left-to-right, right-to-left, top-to-bottom, and bottom-to-top. Each image patch computes the compressed hidden state along the corresponding scan path capturing global context information.