Research Analyzer
← Back ICRA 2026

Density-Aware Point Cloud Upsampling Via Relational Graph Flow Matching

and Shicai Fan

PDF

AI summary

Key figure (auto-extracted from paper)
PURF leverages density-aware relational graphs and transformer-based flow matching to efficiently upsample non-uniform point clouds with superior geometric fidelity and fewer inference steps than diffusion models.
Point cloud upsampling Flow matching Relational graphs Density-aware processing 3D perception Real-time inference

Problem

Conventional point cloud upsampling methods treat all regions homogeneously, causing over-smoothing in sparse areas and redundant computation in dense ones, while diffusion-based alternatives are too computationally heavy for real-time use.

Approach

PURF partitions point clouds into dense and sparse regions using an adaptive relational graph, then employs a transformer-based flow matching model to predict velocity fields for rapid, few-step upsampling.

Key results

  • Achieves state-of-the-art or competitive Chamfer and Hausdorff distances on PUGAN and PU1K benchmarks
  • Demonstrates robust generalization across varying sensor distances in KITTI and different spatial scales in a proprietary campus dataset
  • Reduces inference steps and computational overhead compared to iterative diffusion-based upsampling methods
  • Preserves fine geometric details and minimizes outliers in sparse regions where homogeneous methods fail

Why it matters

Provides a computationally efficient and geometrically accurate upsampling solution critical for real-time 3D perception in autonomous driving and robotics.

Abstract

Real-world point clouds exhibit non-uniform density distributions, varying across distance and scale. Conventional up- sampling methods typically treat points homogeneously, which over-smooths sparse regions while over-processing dense regions. We propose PURF, a density-aware point cloud upsampling frame- work based on relational graph flow matching. PURF leverages a heterogeneous graph representation to capture density variations through relational graph construction, and employs transformer- based flow matching to predict timestep-dependent velocity fields. This design enables a density-aware and efficient mapping from sparse inputs to dense point clouds, reducing computational over- head compared to recent approaches. Extensive experiments on synthetic, KITTI, and a proprietary campus dataset collected by our team demonstrate that PURF achieves advanced performance in upsampling point clouds qualitatively and quantitatively.

Index terms

Deep Learning for Visual Perception Computer Vision for Automation Representation Learning

Related papers