[Paper Review] R$^3$D: Regional-guided Residual Radar Diffusion

Hao Li, Xinqi Liu+1 more|arXiv (Cornell University)|Jan 10, 2026

Advanced SAR Imaging Techniques0 citations

TL;DR

R3D introduces a residual-diffusion framework with sigma-adaptive regional guidance to enhance mmWave radar point clouds by modeling LiDAR-radar residuals and focusing refinement on high-interest regions without extra training costs.

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

Motivation & Objective

• Address sparse and noisy mmWave radar point clouds by improving perception under adverse conditions.
• Reduce learning difficulty by modeling the LiDAR-radar residual instead of the full LiDAR distribution.
• Introduce sigma-adaptive regional guidance to prioritize critical regions during low-noise detail recovery.
• Leverage radar-specific signal properties to generate attention maps without pre-training.
• Demonstrate superior performance on ColoRadar dataset against state-of-the-art methods.

Proposed method

• Model the residual distribution y - x between LiDAR and radar, and train a diffusion model to learn p(r|x).
• Use an exponential noise schedule and denoise the residual conditioned on radar input x and noise level t.
• Construct a radar attention map from intensity and local consistency to identify strong-scattering regions.
• Incorporate an AdaGN-based conditioning in a UNet denoiser with sinusoidal time embeddings.
• Apply sigma-adaptive regional guidance by modifying Karras weights at low-noise stages to emphasize target regions while preserving stability at high-noise stages.
• Infer enhanced radar by iteratively denoising the residual and fusing it with the radar input to produce ŷ = x + r̂.

Experimental results

Research questions

• RQ1 Can modeling the LiDAR-radar residual improve diffusion-based radar enhancement compared to modeling the full LiDAR distribution?
• RQ2 Does sigma-adaptive regional guidance effectively prioritize critical regions without destabilizing training or increasing inference cost?
• RQ3 How well does R3D perform on ColoRadar across diverse scenes relative to state-of-the-art radar enhancement methods?
• RQ4 Can region-aware guidance be achieved without pre-trained domain-adapted features or heavy masking strategies during inference?

Key findings

• Residual diffusion (learning y − x) reduces learning complexity and improves performance over direct LiDAR generation from radar.
• R3D with sigma-adaptive regional guidance outperforms EDM across multiple metrics on ColoRadar (average gains in CD and HD reported in text).
• Regional guidance focused on high-intensity and high-consistency radar regions yields better detail recovery without harming global structure learning.
• DINO-based conditional injection and time-dependent masking can cause domain mismatch or instability, supporting the paper’s principled regional guidance design.
• No additional training or inference costs are incurred compared to strong baselines, while achieving superior results.