Unsupervised Radar Point Cloud Enhancement via Arbitrary LiDAR Guided Diffusion Prior

Yanlong Yang, Jianan Liu, Guanxiong Luo, Hao Li, Euijoon Ahn, Mostafa Rahimi Azghadi, Tao HuangSubmission to ICLR 2026 · Rejected

Abstract

The angular resolution of automotive radar is fundamentally limited by the Rayleigh criterion and the number of antennas, resulting in sparse radar point clouds that hinder high-precision perception. Recent approaches address this limitation by training discriminative or conditional generative models on paired radar-LiDAR data, learning an explicit mapping from radar measurements to LiDAR-like outputs. However, these methods are sensitive to calibration errors and fail when LiDAR data is unavailable. In this work, we propose the first *cross-modality, unsupervised* radar point cloud super-resolution method by formulating radar enhancement as a Bayesian inverse problem. We explicitly model the radar sensing process with a differentiable forward operator, defining the measurement likelihood, and combine it with a LiDAR-trained latent diffusion model that captures the distribution of physically plausible point clouds. During inference, we perform posterior sampling that integrates this generative prior with the radar likelihood, yielding reconstructions that are both geometrically rich and strictly measurement-consistent — without requiring paired radar-LiDAR data. Experiments on the RADIal dataset show that our approach matches the performance of supervised mapping-based methods, while generalizing significantly better to unseen datasets such as K-Radar. These results demonstrate the effectiveness of combining physics-based modeling with distributional priors, offering a robust and practical solution for radar perception in real-world deployments. Our code is available at https://anonymous.4open.science/r/RadarINV2025.
BibTeX
@misc{yang2025unsupervised,
  author = {Yanlong Yang and Jianan Liu and Guanxiong Luo and Hao Li and Euijoon Ahn and Mostafa Rahimi Azghadi and Tao Huang},
  title = {{Unsupervised Radar Point Cloud Enhancement via Arbitrary LiDAR Guided Diffusion Prior}},
  year = {2025},
  eprint = {2505.09887},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2505.09887},
}
Figure