R$^3$D: Regional-guided Residual Radar Diffusion

TL;DR

The paper investigates mmWave radar point cloud enhancement for autonomous perception, targeting sparse, noisy data and high learning complexity in diffusion-based methods. It introduces R$^3$D, a regional-guided residual diffusion framework that models the LiDAR–radar residual and applies $\sigma$-adaptive regional guidance to prioritize refinement in high-value regions while preserving stable training. Across ColoRadar, R$^3$D outperforms direct LiDAR-generation diffusion approaches and prior baselines, with notable improvements in Chamfer Distance and Hausdorff Distance, validating the effectiveness of residual learning and region-aware guidance. The approach generalizes to LiDAR super-resolution tasks and offers a practical solution that does not incur additional training or inference costs, with code and pretrained models released for public use.

Abstract

Millimeter-wave radar enables robust environment perception in autonomous systems under adverse conditions yet suffers from sparse, noisy point clouds with low angular resolution. Existing diffusion-based radar enhancement methods either incur high learning complexity by modeling full LiDAR distributions or fail to prioritize critical structures due to uniform regional processing. To address these issues, we propose R3D, a regional-guided residual radar diffusion framework that integrates residual diffusion modeling-focusing on the concentrated LiDAR-radar residual encoding complementary high-frequency details to reduce learning difficulty-and sigma-adaptive regional guidance-leveraging radar-specific signal properties to generate attention maps and applying lightweight guidance only in low-noise stages to avoid gradient imbalance while refining key regions. Extensive experiments on the ColoRadar dataset demonstrate that R3D outperforms state-of-the-art methods, providing a practical solution for radar perception enhancement. Our anonymous code and pretrained models are released here: https://anonymous.4open.science/r/r3d-F836

Figures (4)

• Figure 1: Residual Learning Concentrates Target Distribution Near Zero. Left: High-Res LiDAR target (Active: 2.6%, Range: 255, full value distribution); Right: Residual (LiDAR - Radar) target (Active: 6.8%, Range: 338 (absolute range), Std: 15.6, centered near zero). Although the residual has a larger absolute range, 93.2% of its valid values are within ±10 (highly concentrated around zero), resulting in a significantly smaller standard deviation. This concentration makes residual learning substantially easier than modeling the full LiDAR distribution, as the model only needs to predict small deviations from zero rather than the entire value spectrum.
• Figure 2: mmWave radar signal processing pipeline.
• Figure 3: Overview of the R$^3$D framework. Our method integrates residual learning with $\sigma$-adaptive regional guidance for radar point cloud enhancement. The pipeline consists of: (1) Data preprocessing converting LiDAR and radar point clouds to BEV images, (2) Residual computation and forward diffusion with exponential noise schedule, (3) Attention map generation from radar signal properties, (4) UNet denoiser with radar conditioning and noise level embedding, (5) $\sigma$-adaptive guidance module applying region-specific weights only at low noise levels, and (6) Residual fusion producing the enhanced point cloud output.
• Figure 4: Performance comparison on KITTI-360 LiDAR super-resolution.