PVNet: Point-Voxel Interaction LiDAR Scene Upsampling via Diffusion Models.

Area of Science:

• Computer Vision
• 3D Scene Understanding
• LiDAR Technology

Background:

• High-quality point clouds are crucial for 3D scene understanding in outdoor environments.
• LiDAR data sparsity severely limits downstream 3D perception tasks.
• Existing upsampling methods lack generalization for complex outdoor scenes.

Purpose of the Study:

• To introduce PVNet, a novel framework for LiDAR point cloud upsampling.
• To address the limitations of object-centric upsampling methods in complex scenes.
• To enable scene-level point cloud upsampling without dense supervision.

Main Methods:

• Utilizing a diffusion model (DDPMs) with classifier-free guidance for point cloud generation.
• Employing sparse point clouds as guiding conditions and synthesizing data from nearby frames.
• Designing a voxel completion module for feature refinement and enrichment.
• Integrating point and voxel features through a dedicated interaction module.

Main Results:

• PVNet achieves state-of-the-art performance on various benchmarks.
• The method demonstrates effective upsampling for complex outdoor scenes.
• It supports arbitrary upsampling rates, a novel capability for scene-level methods.

Conclusions:

• PVNet offers a robust solution for LiDAR point cloud upsampling in outdoor scenes.
• The point-voxel interaction framework enhances environmental perception.
• This work presents the first scene-level upsampling method with flexible rate support.