Research map / 3D vision / Gaussian splatting
3D Point Splatting for mmWave Radar Novel View Synthesis
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes 3D Point Splatting (3DPS), a differentiable point-based radar renderer with ITU-R P.2040 material BSDFs and PSF splatting of complex range profiles, enabling ADC, CRP and RA outputs from one optimized scene.
Problem
Novel view synthesis for mmWave FMCW radar needs a renderer that is physically faithful, complex-valued, and tractable across many training views; existing MC ray tracers are slow and optical-NVS adaptations discard phase and use opaque features.
Approach
Derive a closed-form, deterministic radar renderer from the solid-angle form of the radar equation: oriented 3D points with ITU-R P.2040 material parameters contribute complex phasors that are splatted into range bins through a precomputed Hann-windowed PSF, with fused CUDA kernels.
Key results
- 3DPS reaches 0.587 mean test Pearson correlation on held-out |RA| across six ColoRadar scenes, versus 0.339 for RadarSplat, 0.133 for Radar Fields, and 0.112 for DART.
- On training views 3DPS reaches 0.812 mean |RA| correlation and 0.701 CRP correlation (test 0.603).
- Training takes about 3 minutes per scene (8 viewpoints, 500 iterations) on a single RTX 4090, versus 19.7 min for RadarSplat.
Surprising findings
- Radar Fields and DART achieve higher PSNR or SSIM than expected relative to their correlation scores on some metrics, showing metric disagreement among baselines.
- Test-train gap of 0.225 in Pearson correlation for 3DPS, attributed to cascading sparsity and quasi-mirror specular returns.
Limitations
- Single-bounce path model; does not handle indoor multipath.
- Requires co-registered LiDAR for initialization; LiDAR-free radar NVS left to future work.
- Train-test gap of 0.225 in correlation indicates limited generalization to held-out views.
Experimental setup
Datasets: ColoRadar
Benchmarks: ColoRadar
Metrics: Pearson correlation PSNR SSIM RMSE relative phase coherence ADC envelope correlation
Baselines: DART Radar Fields RadarSplat
Builds on: 3D Gaussian Splatting Radar Fields RadarSplat Sionna RT ITU-R P.2040
Closest papers by meaning
- RxGS: Receiver-Generalizable 3D Gaussian Splatting for Radio-Frequency Data Synthesis (NeurIPS 2026 · similarity 0.84)
- HGSFusion: Radar-Camera Fusion with Hybrid Generation and Synchronization for 3D Object Detection (AAAI 2025 · similarity 0.83)
- DART: Implicit Doppler Tomography for Radar Novel View Synthesis (CVPR 2024 · similarity 0.83)
- SplatFormer: Point Transformer for Robust 3D Gaussian Splatting (ICLR 2025 · similarity 0.83)
- Subsurface Scattering for Gaussian Splatting (NeurIPS 2024 · similarity 0.83)
- GS-LiDAR: Generating Realistic LiDAR Point Clouds with Panoramic Gaussian Splatting (ICLR 2025 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.