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

Area of Science:

• Computer Vision
• Robotics
• Machine Learning

Background:

• LiDAR upsampling is crucial for generating high-resolution 3D point clouds from sparse inputs, vital for autonomous driving and robotics perception.
• Current methods often convert 3D LiDAR data to 2D range images and use convolutional or Transformer architectures for super-resolution.
• Adapting advanced sequence models like State Space Models (SSMs) to LiDAR upsampling presents a promising research avenue.

Purpose of the Study:

• To introduce LU-Mamba, a novel framework for LiDAR upsampling utilizing the Mamba architecture.
• To tailor the Mamba model specifically for the directional characteristics of LiDAR range images, focusing on vertical axis upsampling.
• To improve the accuracy and efficiency of reconstructing dense point clouds from sparse LiDAR data.

Main Methods:

• Developed LU-Mamba, a framework based on Mamba, an efficient State Space Model (SSM) for long sequence modeling.
• Introduced a Bidirectional Mamba Group to capture long-range vertical dependencies by scanning image rows top-down and bottom-up.
• Incorporated a Direction-Aware Mamba Module with directional modulation to enhance sensitivity to scanning direction and adaptive fusion.

Main Results:

• LU-Mamba achieved state-of-the-art performance on benchmark LiDAR upsampling datasets.
• The proposed method surpassed existing convolutional and Transformer-based approaches in point cloud reconstruction quality.
• Ablation studies confirmed the effectiveness of the Bidirectional Mamba Group and Direction-Aware Mamba Module.

Conclusions:

• LU-Mamba represents a significant advancement in LiDAR upsampling by leveraging Mamba's sequence modeling capabilities.
• The directional adaptations are key to effectively processing LiDAR range images for improved perception tasks.
• The framework offers a new, high-performing alternative for high-resolution 3D point cloud generation in autonomous systems.