LU-mamba: LiDAR upsampling via bidirectional state space modeling on range images.

Abstract

LiDAR upsampling seeks to reconstruct dense, high-resolution point clouds from sparse LiDAR inputs-an essential task for enhancing perception in autonomous driving and robotics. Although inherently 3D, this problem is typically reformulated as a 2D range image super-resolution task by projecting point clouds onto a spherical range image. While existing methods predominantly rely on convolutional or Transformer-based architectures, adapting emerging sequence models to this task remains an open and promising direction. In this paper, we introduce LU-Mamba, a novel LiDAR upsampling framework built upon Mamba, a state space model (SSM) architecture designed for efficient long-sequence modeling. Unlike general-purpose vision models, LU-Mamba is tailored to the directional structure of LiDAR range images, where upsampling is performed strictly along the vertical axis. To this end, LU-Mamba introduces two key innovations: (1) a Bidirectional Mamba Group, which captures long-range vertical dependencies by repeatedly scanning image row sequences in both top-down and bottom-up directions; and (2) a Direction-Aware Mamba Module that explicitly encodes the scanning direction with a directional modulation mechanism, enhancing directional sensitivity and adaptive fusion. Extensive experiments on benchmark LiDAR upsampling datasets demonstrate that LU-Mamba achieves new state-of-the-art performance, surpassing both convolutional and Transformer-based baselines. Ablation studies further validate the effectiveness of our architectural designs.