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

2026-07-27 · Source: cs.CV updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Expert, extended

Summary

JustDepth is a novel single-stage radar-camera depth estimator designed for autonomous systems, trained exclusively with radar, camera, and single-scan LiDAR data, avoiding auxiliary annotations. It aggregates all radar returns into a fixed-width 1D representation, ensuring constant inference latency regardless of point count. The architecture incorporates a Height Fusion Block for efficient modality fusion and a lightweight Graph Neural Network (GNN) for global depth propagation. A training-only confidence decoder stabilizes learning without adding test-time overhead. JustDepth effectively mitigates LiDAR Distribution Leakage (LDL) stripe artifacts through point upsampling and synchronized rotation augmentation, quantified by the Vertical-Horizontal Gradient Ratio (VHGR). On the nuScenes dataset, it achieves real-time inference at approximately 14.8 ms, reducing inference time by 39.7x and stripe artifacts by 66% compared to recent state-of-the-art methods, while maintaining competitive accuracy, including the best AbsRel score of 0.074.

Key takeaway

For Machine Learning Engineers developing real-time depth estimation for autonomous driving, JustDepth offers a compelling solution. Its single-stage, constant-latency architecture, which processes frames in approximately 14.8 ms, significantly reduces inference time by 39.7x while mitigating LiDAR Distribution Leakage artifacts by 66%. You should consider integrating this approach to simplify your perception stack, reduce reliance on auxiliary annotations, and achieve robust, artifact-free depth maps crucial for safety-critical applications.

Key insights

Single-stage radar-camera depth estimation achieves real-time performance and artifact reduction with minimal supervision.

Principles

• Fixed-width radar encoding ensures constant inference latency.
• GNNs effectively propagate depth cues globally for dense prediction.
• Training-only auxiliary decoders improve accuracy without test-time cost.

Method

Fuses 1D radar and 2D image features via a Height Fusion Block, then propagates depth globally using a GNN, supervised by single-scan LiDAR with augmentations.

In practice

• Use point upsampling and synchronized rotation for LDL mitigation.
• Quantify stripe artifacts with Vertical-Horizontal Gradient Ratio (VHGR).
• Employ a training-only confidence decoder for accuracy gains.

Topics

• Radar-Camera Fusion
• Depth Estimation
• Autonomous Driving
• Graph Neural Networks
• Real-time Inference
• LiDAR Distribution Leakage

Code references

• TPyun/JustDepth

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, Computer Vision Engineer

Related on AIssential

• JustDepth: Real-Time Radar-Camera Depth Estimation with Single-Scan LiDAR Supervision — JustDepth integrates radar and camera for real-time depth estimation, using single-scan LiDAR for supervision.
• LIDAR-AD: A Decoder-Free Latent-Interaction Dreamer with Action-Residual Chains for Autonomous Driving — LIDAR-AD improves autonomous driving by focusing on risk-relevant latent interactions and residual action modeling for long-horizon control.
• Towards Robust Driving Perception: A Flexible Scale-Driven Family for Self-Supervised Monocular Depth Estimation — FlexDepth improves monocular depth estimation for driving by decoupling static/dynamic training and using a scale-driven decoder.
• 🔥ZipDepth: Depth on Any Device🔥 👉ZipDepth from UniBO is a super-compact monocular depth... — ZipDepth achieves compact monocular depth estimation via reparameterizable encoder-decoder and knowledge distillation from a foundation model.
• Large Depth Completion Model from Sparse Observations — LDCM achieves robust, metric-accurate dense depth completion from sparse observations via a transformer, novel initialization, and 3D coordinate regression.
• A Comprehensive Survey on Deep Learning-Based LiDAR Super-Resolution for Autonomous Driving — LiDAR super-resolution uses deep learning to enhance sparse point clouds, enabling cheaper sensors to perform like high-resolution ones.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CV updates on arXiv.org.