High-Fidelity 3D Reconstruction From Sparse Multi-Baseline CSAR Data via Conditional Diffusion Models

Abstract

Three-dimensional (3D) reconstruction from sparse multibaseline circular synthetic aperture radar (CSAR) data are a significant challenge, often plagued by low vertical resolution, geometric blurring and severe artefacts due to limited observation angles. To address this, this paper introduces a 3D reconstruction framework based on a conditional diffusion model. The proposed method utilises a dual-path 3D U-Net, where a condition encoder extracts geometric features from a coarse initial reconstruction. These features guide the progressive denoising process via cross-attention, integrating physical constraints with data-driven generation. Furthermore, the framework incorporates a novel cross-view consistency loss to enforce geometric coherence across sparse views, enhancing structural integrity. Experiments demonstrate superior performance in quantitative metrics (IoU and BCE) and visual quality, successfully restoring complete and detailed 3D structures.