Research map / 3D vision / NeRF and radiance fields
DART: Implicit Doppler Tomography for 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 DART, a NeRF-inspired method that learns an implicit reflectance/transmittance field from range-Doppler radar scans to render novel-view radar images and produce tomographic maps.
Problem
Simulating realistic mmWave radar scans requires accurate scene geometry and radar-specific material properties, which are costly to specify manually; compact radars also have poor angular resolution.
Approach
Adapts NeRF to radar by rendering range-Doppler pixels as integrals over velocity-dependent rings (circles) rather than rays, with view-dependent reflectance and transmittance modeled by Instant NGP hash grids and spherical harmonics, trained by gradient descent on handheld radar traces.
Key results
- DART achieves mean SSIM 0.636 on the test sets, versus 0.545 for CFAR, 0.468 for Nearest Neighbor, and 0.463 for the Lidar simulator (Table 1).
- Removing view dependence drops mean SSIM to 0.614 and using a 20cm grid drops it to 0.591, showing both components matter.
- DART trains for 3 epochs per dataset on one RTX 4090 in roughly 1-2x the data collection time (~10 minutes).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace Lidar with DART | Radar novel view synthesis · All datasets | Mean SSIM: 0.174 ± 0.013 | Helped |
| replace Nearest with DART | Radar novel view synthesis · All datasets | Mean SSIM: 0.168 ± 0.012 | Helped |
| replace CFAR with DART | Radar novel view synthesis · All datasets | Mean SSIM: 0.091 ± 0.006 | Helped |
| add view dependence using spherical harmonics (ablation) | DART · All datasets | Mean SSIM: 0.022 ± 0.005 | Helped |
| use a dynamic grid representation instead of a 20cm grid (ablation) | DART · All datasets | Mean SSIM: 0.046 ± 0.004 | Helped |
Surprising findings
- Lidar-based simulation with a fixed reflectance model is worse than the nearest-neighbor baseline built from real radar data, despite lidar giving accurate geometry.
Limitations
- Assumes a static scene and a constantly moving radar with accurate velocity estimates.
- Requires accurate ground-truth pose, obtained from lidar-based SLAM.
- Lower spatial resolution than lidar; cannot resolve small weak reflectors well.
Experimental setup
Datasets: custom handheld mmWave radar traces with lidar-based pose and velocity
Benchmarks: custom DART radar dataset (12 traces)
Metrics: SSIM standard error of mean SSIM SSIM of Gaussian-noise reference levels
Baselines: Lidar scan-based simulator Nearest Neighbor CFAR point cloud aggregation
Builds on: NeRF Instant NGP Bartlett direction-of-arrival estimation Cartographer
Closest papers by meaning
- 3D Point Splatting for mmWave Radar Novel View Synthesis (NeurIPS 2026 · similarity 0.83)
- TransLoc4D: Transformer-based 4D Radar Place Recognition (CVPR 2024 · similarity 0.82)
- NEAR: Neural Electromagnetic Array Response (ICML 2025 · similarity 0.82)
- Selection, Not Fusion: Radar-Modulated State Space Models for Radar-Camera Depth Estimation (NeurIPS 2026 · similarity 0.81)
- RadarMOSEVE: A Spatial-Temporal Transformer Network for Radar-Only Moving Object Segmentation and Ego-Velocity Estimation (AAAI 2024 · similarity 0.81)
- LiDAR-RT: Gaussian-based Ray Tracing for Dynamic LiDAR Re-simulation (CVPR 2025 · similarity 0.80)
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.