SpINR: Neural Volumetric Reconstruction for FMCW Radars

TL;DR

SpINR tackles the challenge of high-fidelity 3D reconstruction from FMCW radar by marrying a closed-form, differentiable forward model in the frequency domain with implicit neural representations for continuous volumetric scenes. It exploits the linear beat-frequency relationship $f_b = S\tau$ to map distances to frequency bins, enabling efficient supervision using only the relevant bins and bypassing costly time-domain simulations. The framework demonstrates state-of-the-art reconstruction quality across multiple objects and metrics, outperforming classical backprojection and time-domain learning baselines while offering greater stability and computational efficiency. This work establishes the first end-to-end neural volumetric reconstruction pipeline for radar data, with strong implications for radar-based perception systems and future multi-sensor fusion.

Abstract

In this paper, we introduce SpINR, a novel framework for volumetric reconstruction using Frequency-Modulated Continuous-Wave (FMCW) radar data. Traditional radar imaging techniques, such as backprojection, often assume ideal signal models and require dense aperture sampling, leading to limitations in resolution and generalization. To address these challenges, SpINR integrates a fully differentiable forward model that operates natively in the frequency domain with implicit neural representations (INRs). This integration leverages the linear relationship between beat frequency and scatterer distance inherent in FMCW radar systems, facilitating more efficient and accurate learning of scene geometry. Additionally, by computing outputs for only the relevant frequency bins, our forward model achieves greater computational efficiency compared to time-domain approaches that process the entire signal before transformation. Through extensive experiments, we demonstrate that SpINR significantly outperforms classical backprojection methods and existing learning-based approaches, achieving higher resolution and more accurate reconstructions of complex scenes. This work represents the first application of neural volumetic reconstruction in the radar domain, offering a promising direction for future research in radar-based imaging and perception systems.

Figures (7)

• Figure 1: Overview of our method. FMCW radar measurements are collected from multiple viewpoints and transformed to the frequency domain for supervision. An implicit neural representation (INR) predicts scatterer intensity over the scene, and our differentiable spectral forward model synthesizes the complex-valued frequency-domain signal (magnitude and phase). The INR is trained by minimizing the discrepancy between synthesized and measured frequency-domain signals, using only the bins corresponding to the valid scene distances.
• Figure 2: When the $\alpha = \beta_k$, there is no leakage
• Figure 3: When the $\alpha \neq \beta_k$, there is leakage and in neighbouring bins
• Figure 4: Comparison of volumetric reconstructions for Range Quantization, Time-domain forward model with Temporal Supervision, Time-domain forward model with Spectral Supervision, and our method. Our reconstructions more accurately match the ground truth geometry compared to other methods.
• Figure 5: We measure the runtime of different forward models with different scene size governed by the number of scatterers. Our proposed method outperforms the time domain forward model while range quantization is the fastest.