3D Point Splatting for mmWave Radar Novel View Synthesis

TL;DR

3D Point Splatting (3DPS) introduces a differentiable point-based renderer for mmWave radar that accurately models complex-valued radar signals using physical principles. It enables multi-view novel view synthesis with high fidelity across ADC, complex range profile, and range-azimuth outputs without retraining. The method achieves state-of-the-art performance on outdoor radar scenes with fast training times.

Why It Matters

This work enables accurate, physics-driven simulation and synthesis of mmWave radar data, which is critical for training robust perception systems in autonomous driving and robotics. By supporting multiple output formats from a single trained model, it reduces the need for task-specific retraining and accelerates development cycles.

Key Findings

• 3DPS is the first differentiable point renderer for radar that uses a closed-form material model based on ITU-R P.2040 and a precomputed point spread function.
• The renderer produces complex-valued outputs that support multiple downstream formats—ADC, CRP, and RA—via standard FFT pipelines without retraining.
• On ColoRadar datasets, 3DPS achieves a mean Pearson correlation of 0.587 on unseen RA views, outperforming optical-NVS baselines by 1.7x to 5.2x.