JustDepth: Real-Time Radar-Camera Depth Estimation with Single-Scan LiDAR Supervision

2026-07-24

AI Summary

This work addresses the high latency and poor portability of existing radar-camera depth estimation methods, which typically rely on multi-stage pipelines and auxiliary annotations. To overcome these limitations, we propose the first end-to-end, single-stage framework that requires only single-frame LiDAR supervision. Our approach aggregates radar echoes into a fixed-width one-dimensional representation and leverages a height-aware fusion module combined with a lightweight graph neural network to enable efficient global depth propagation. A confidence decoder is introduced during training—imposing zero overhead at inference—to stabilize learning, while a novel vertical-horizontal gradient ratio (VHGR) metric is designed to quantify and substantially suppress stripe-like artifacts. On nuScenes, our method achieves a 39.7× speedup in inference over the current state of the art, reduces stripe artifacts by 66% (measured by VHGR), and maintains competitive accuracy.

Abstract

Accurate yet low-latency depth is essential for radar-camera perception in autonomous systems. Cameras provide rich appearance but lack metric scale, whereas automotive radar offers metric range but is sparse and noisy. Many pipelines are multi-stage or depend on auxiliary annotations, increasing latency and limiting portability. We introduce JustDepth, a single-stage radar-camera depth estimator trained only with radar, camera, and single-scan LiDAR. All radar returns are aggregated into a fixed-width 1D representation, decoupling runtime from point count. A Height Fusion Block fuses modalities, a lightweight GNN propagates depth globally, and a training-only confidence decoder stabilizes learning with zero test-time cost. We mitigate stripe artifacts via simple augmentations and quantify them using the Vertical-Horizontal Gradient Ratio (VHGR). On nuScenes, compared to recent state-of-the-art methods, JustDepth maintains accuracy while reducing inference time by 39.7x and stripe artifacts by 66% as measured by VHGR.

Problem

• depth estimation
• radar-camera fusion
• low-latency perception
• autonomous driving
• LiDAR supervision

Innovation

• radar-camera fusion
• single-stage depth estimation
• LiDAR supervision
• stripe artifact reduction
• real-time inference