3D Point Splatting for mmWave Radar Novel View Synthesis
The paper proposes 3DPS as a radar-native renderer that keeps phase, material modeling, and multi-view training in one method.
3D Point Splatting derives its renderer from the radar equation and assigns each oriented 3D point an ITU-R P.2040 material model. Its complex-valued output can produce ADC, CRP, and RA formats through FFT pipelines without retraining. On six outdoor ColoRadar scenes, it reports 0.587 mean Pearson correlation on held-out RA images, ahead of three optical-NVS baselines. Training is listed at about 3 minutes per scene on a single RTX 4090. ArXiv · AI/CL/LG's note
3D Point Splatting derives its renderer from the radar equation and assigns each oriented 3D point an ITU-R P.2040 material model. Its complex-valued output can produce ADC, CRP, and RA formats through FFT pipelines without retraining. On six outdoor ColoRadar scenes, it reports 0.587 mean Pearson correlation on held-out RA images, ahead of three optical-NVS baselines. Training is listed at about 3 minutes per scene on a single RTX 4090. ArXiv · AI/CL/LG's note
score 4