International Conferences
| LiDAR-to-4D Radar Synthesis for Building Large-Scale Tensor Datasets |
| Year |
2026 |
| Month |
|
| Journal |
2026 / [Top AI Conference] IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR-Findings) |
| Author |
Woo-Jin Jung, Dong-Hee Paek, and Seung-Hyun Kong* |
|
4-dimensional (4D) radar tensors preserve dense spatial measurements, offering strong potential to improve object detection robustness in both adverse and normal driving conditions. However, publicly available autonomous driving datasets containing 4D radar tensors are significantly fewer and less diverse than LiDAR datasets, which limits the generalization capability of perception models. To address this limitation, we propose LiDAR-to-4D radar data synthesis (L2RDaS), a framework that expands the scale and diversity of 4D radar tensor datasets by synthesizing Cartesian range–azimuth–elevation (C-RAE) tensors from LiDAR data. L2RDaS supports two data augmentation modes: (1) dataset expansion, which synthesizes tensors from datasets without 4D radar tensors, and (2) 4D radar tensor ground-truth augmentation (GT-Aug), which synthesizes tensors that reflect real radar distributions, including sidelobes. Experiments on the K-Radar dataset demonstrate that L2RDaS significantly improves bird’s-eye-view (BEV) and 3D object detection performance, validating its effectiveness as a data augmentation framework for 4D radar-based perception. All codes and logs will be available after the review. |
