검색 상세

BlitzDepth : Real-Time Radar-Camera Depth Estimation Using a Confidence Branch and Global GNN Refinement

초록/요약

Accurate depth estimation with low inference latency is essential for radar–camera per- ception in autonomous systems. While cameras supply dense appearance cues but no metric scale, automotive radar provides metrically valid ranges that are sparse and noisy. Existing approaches are often built as multi-stage pipelines or rely on extra labels such as semantic or panoptic annotations, which increase runtime and reduce flexibility across datasets and sensor setups. We introduce BlitzDepth, a single-stage radar–camera depth estimation network trained using only radar, RGB images, and single-scan LiDAR supervision. All radar echoes from a frame are compressed into a fixed-width one-dimensional representation so that the compu- tational cost is independent of the number of radar points. A Height Fusion Block aligns and combines radar and image features, a compact graph neural network (GNN) spreads depth information across the scene, and an auxiliary confidence decoder used only during training stabilizes optimization without any test-time overhead. To address stripe artifacts caused by LiDAR scanlines, we incorporate simple data augmentations and evaluate the resulting Li- DAR Distribution Leakage using the Vertical–Horizontal Gradient Ratio (VHGR). On the nuScenes benchmark, BlitzDepth achieves depth accuracy comparable to recent state-of-the- art methods while reducing inference time by 39.7× and lowering stripe artifacts by 66% according to VHGR.

more

목차

1 Introduction 1
2 Related works 4
3 Methods 6
A Fusion Method 7
B Graph-Based Depth Propagation 8
C Depth Decoder 10
D Confidence Decoder (Training Only) 11
E LiDAR Distribution Leakage Mitigation 12
4 Experiments 14
A Dataset and Evaluation Metrics 14
B Implementation Details 14
C Quantitative Results 16
D Qualitative Results 16
E Gradient Metrics 18
F Ablation Study 19
5 Conclusion 25
References 27
국문 초록 31

more
반출 Meta View 목록