Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

XD-RCDepth: Lightweight Radar-Camera Depth Estimation with Explainability-Aligned and Distribution-Aware Distillation

Huawei Sun, Zixu Wang, Xiangyuan Peng, Julius Ott, Georg Stettinger, Lorenzo Servadei, Robert Wille

TL;DR

XD-RCDepth introduces a lightweight radar–camera depth estimation framework designed for autonomous driving, achieving significant parameter reduction while preserving accuracy. The approach combines a compact FiLM-based fusion with a Point-wise DASPP decoder, and couples this with two distillation strategies: X-KD transfers the teacher’s saliency structure to the student, and D2-KD recasts depth regression as soft bin classification to align depth distributions. Empirical results on nuScenes and ZJU-4DRadarCam show the model reaches around 8.9M parameters (about 30% fewer than LiRCDepth) and real-time performance (~15 FPS), with consistent MAE improvements of up to about 8% at longer ranges when using both distillation methods. These contributions enhance robustness under compression and improve interpretability, making the method practical for real-world autonomous driving in adverse conditions.

Abstract

Depth estimation remains central to autonomous driving, and radar-camera fusion offers robustness in adverse conditions by providing complementary geometric cues. In this paper, we present XD-RCDepth, a lightweight architecture that reduces the parameters by 29.7% relative to the state-of-the-art lightweight baseline while maintaining comparable accuracy. To preserve performance under compression and enhance interpretability, we introduce two knowledge-distillation strategies: an explainability-aligned distillation that transfers the teacher's saliency structure to the student, and a depth-distribution distillation that recasts depth regression as soft classification over discretized bins. Together, these components reduce the MAE compared with direct training with 7.97% and deliver competitive accuracy with real-time efficiency on nuScenes and ZJU-4DRadarCam datasets.

XD-RCDepth: Lightweight Radar-Camera Depth Estimation with Explainability-Aligned and Distribution-Aware Distillation

TL;DR

XD-RCDepth introduces a lightweight radar–camera depth estimation framework designed for autonomous driving, achieving significant parameter reduction while preserving accuracy. The approach combines a compact FiLM-based fusion with a Point-wise DASPP decoder, and couples this with two distillation strategies: X-KD transfers the teacher’s saliency structure to the student, and D2-KD recasts depth regression as soft bin classification to align depth distributions. Empirical results on nuScenes and ZJU-4DRadarCam show the model reaches around 8.9M parameters (about 30% fewer than LiRCDepth) and real-time performance (~15 FPS), with consistent MAE improvements of up to about 8% at longer ranges when using both distillation methods. These contributions enhance robustness under compression and improve interpretability, making the method practical for real-world autonomous driving in adverse conditions.

Abstract

Depth estimation remains central to autonomous driving, and radar-camera fusion offers robustness in adverse conditions by providing complementary geometric cues. In this paper, we present XD-RCDepth, a lightweight architecture that reduces the parameters by 29.7% relative to the state-of-the-art lightweight baseline while maintaining comparable accuracy. To preserve performance under compression and enhance interpretability, we introduce two knowledge-distillation strategies: an explainability-aligned distillation that transfers the teacher's saliency structure to the student, and a depth-distribution distillation that recasts depth regression as soft classification over discretized bins. Together, these components reduce the MAE compared with direct training with 7.97% and deliver competitive accuracy with real-time efficiency on nuScenes and ZJU-4DRadarCam datasets.
Paper Structure (13 sections, 10 equations, 3 figures, 4 tables)

This paper contains 13 sections, 10 equations, 3 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Methodology
  3. Model Architecture
  4. Explainability-Aligned Saliency Map Distillation
  5. Depth Distribution Distillation
  6. Loss Functions
  7. Experiments
  8. Datasets and Implementation Details
  9. Quantitative Results
  10. Qualitative Results
  11. Ablation Studies
  12. Conclusion
  13. Acknowledgement

Figures (3)

  • Figure 1: The main structure of our XD-RCDepth, along with the proposed distillation methods.
  • Figure 2: Qualitative comparison between the XD-RCDepth trained with and without distillation.
  • Figure 3: Saliency map comparison between the teacher and the student model trained with and without our distillation.