3D Point Splatting for mmWave Radar Novel View Synthesis

mmWave radar novel view synthesis has been stuck between two bad options: physically grounded Monte Carlo ray tracing that is too slow for optimization, and fast optical-NVS-style methods that throw away phase and explicit material behavior. 3D Point Splatting (3DPS) tries to close that gap with a differentiable point renderer built directly from the solid-angle form of the radar equation.

Each oriented 3D point in the scene carries an ITU-R P.2040 material model, evaluated in closed form, and the complex phasor is splatted into range bins through a precomputed point spread function. That matters because the renderer stays complex-valued, which makes it usable across radar output formats instead of being locked to power-only range-azimuth magnitudes.

The same optimized scene can generate ADC, complex range profile, and range-azimuth outputs through standard FFT pipelines without retraining for each representation. In practice, that gives radar researchers a more flexible scene model for inverse rendering, sensor simulation, and viewpoint synthesis workflows where phase and material response are not optional details.

On six outdoor ColoRadar scenes, 3DPS reached 0.587 mean Pearson correlation on held-out range-azimuth images. The result sits between 1.7x and 5.2x the optical-NVS baselines named RadarSplat, Radar Fields, and DART, while training in roughly 3 minutes per scene on a single RTX 4090.