RadarFuser: One-Stage Radar-Camera Depth Estimation With Early and Point-Guided Fusion

Abstract

Depth estimation is fundamental to vision applications, but recovering metric depth from a single image remains challenging due to scale ambiguity. Radar-camera depth estimation offers a practical solution by combining dense visual cues with weather-robust radar measurements, yet the extreme sparsity and noise of radar returns make cross-modal fusion difficult. Existing methods typically adopt late fusion with either 2D radar maps or 3D radar point clouds, and thus fail to fully exploit the complementary cues from both representations. In addition, LiDAR supervision provides accurate metric scale but remains spatially sparse, leaving the use of monocular depth for dense structural supervision underexplored. To this end, we propose RadarFuser, a novel one-stage radar-camera depth estimation framework with early and point-guided fusion. RadarFuser fuses the image-projected 2D radar map early with RGB to exploit image-plane alignment, and injects complementary 3D radar geometric cues through a point branch and a Radar Cross-Attention Fusion (RCAF) module. To compensate for sparse LiDAR supervision, RadarFuser exploits monocular depth for dense structural supervision, while avoiding direct enforcement of its unreliable absolute scale. Experiments on nuScenes demonstrate that RadarFuser achieves state-of-the-art performance and remains robust under adverse weather, while evaluation on ZJU-4DRadarCam further verifies its effectiveness on denser radar data.