Title: XD-RCDepth: Efficient Radar-Camera Depth Estimation via Explainability-Aligned and Distribution-Aware Distillation

Abstract: Accurate depth estimation is a cornerstone of autonomous driving systems, and the fusion of radar and camera data enhances reliability in challenging environments by leveraging complementary geometric information. This study introduces XD-RCDepth, a streamlined architecture that achieves a 29.7% reduction in parameters compared to the leading lightweight baseline, without compromising on accuracy. To ensure robust performance during model compression and to improve interpretability, we propose two distinct knowledge-distillation techniques. The first, explainability-aligned distillation, transfers the teacher model’s saliency patterns to the student network. The second, depth-distribution distillation, reframes depth regression as a soft classification task across discretized bins. Combined, these methods yield a 7.97% improvement in Mean Absolute Error (MAE) over direct training approaches, offering competitive accuracy and real-time processing speeds on the nuScenes and ZJU-4DRadarCam datasets.

Code: https://github.com/harborsarah/XD_RCDepth

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC