Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

L2RDaS: Synthesizing 4D Radar Tensors for Model Generalization via Dataset Expansion

Woo-Jin Jung, Dong-Hee Paek, Seung-Hyun Kong

TL;DR

This work tackles the scarcity and limited diversity of public 4D radar tensor data by introducing L2RDaS, a framework that synthesizes spatially informative 4D radar tensors from LiDAR data. The generator extends a pix2pixHD-like architecture with 3D voxelized input, sparse-to-dense 3D convolutions, and a multi-scale 3D discriminator, while the OBIS module injects object-level cues (boundary and Gaussian points) to improve reflection fidelity and spatial alignment. Ground-truth augmentation (GT-Aug) is supported by embedding annotated objects into LiDAR and synthesizing them into radar tensors, enabling realistic clutter and object measurements. Experiments on the K-Radar dataset show that dataset expansion with synthesized tensors yields average improvements in ${{AP}_{BEV}}$ and ${{AP}_{3D}}$ across detectors, and GT-Aug provides additional gains, highlighting practical benefits for generalization without extra sensing. The work points to future directions, including incorporating temporal LiDAR to recover Doppler information and further refining alignment with real radar distributions.

Abstract

4-dimensional (4D) radar is increasingly adopted in autonomous driving for perception tasks, owing to its robustness under adverse weather conditions. To better utilize the spatial information inherent in 4D radar data, recent deep learning methods have transitioned from using sparse point cloud to 4D radar tensors. However, the scarcity of publicly available 4D radar tensor datasets limits model generalization across diverse driving scenarios. Previous methods addressed this by synthesizing radar data, but the outputs did not fully exploit the spatial information characteristic of 4D radar. To overcome these limitations, we propose LiDAR-to-4D radar data synthesis (L2RDaS), a framework that synthesizes spatially informative 4D radar tensors from LiDAR data available in existing autonomous driving datasets. L2RDaS integrates a modified U-Net architecture to effectively capture spatial information and an object information supplement (OBIS) module to enhance reflection fidelity. This framework enables the synthesis of radar tensors across diverse driving scenarios without additional sensor deployment or data collection. L2RDaS improves model generalization by expanding real datasets with synthetic radar tensors, achieving an average increase of 4.25\% in ${{AP}_{BEV}}$ and 2.87\% in ${{AP}_{3D}}$ across three detection models. Additionally, L2RDaS supports ground-truth augmentation (GT-Aug) by embedding annotated objects into LiDAR data and synthesizing them into radar tensors, resulting in further average increases of 3.75\% in ${{AP}_{BEV}}$ and 4.03\% in ${{AP}_{3D}}$. The implementation will be available at https://github.com/kaist-avelab/K-Radar.

L2RDaS: Synthesizing 4D Radar Tensors for Model Generalization via Dataset Expansion

TL;DR

This work tackles the scarcity and limited diversity of public 4D radar tensor data by introducing L2RDaS, a framework that synthesizes spatially informative 4D radar tensors from LiDAR data. The generator extends a pix2pixHD-like architecture with 3D voxelized input, sparse-to-dense 3D convolutions, and a multi-scale 3D discriminator, while the OBIS module injects object-level cues (boundary and Gaussian points) to improve reflection fidelity and spatial alignment. Ground-truth augmentation (GT-Aug) is supported by embedding annotated objects into LiDAR and synthesizing them into radar tensors, enabling realistic clutter and object measurements. Experiments on the K-Radar dataset show that dataset expansion with synthesized tensors yields average improvements in and across detectors, and GT-Aug provides additional gains, highlighting practical benefits for generalization without extra sensing. The work points to future directions, including incorporating temporal LiDAR to recover Doppler information and further refining alignment with real radar distributions.

Abstract

4-dimensional (4D) radar is increasingly adopted in autonomous driving for perception tasks, owing to its robustness under adverse weather conditions. To better utilize the spatial information inherent in 4D radar data, recent deep learning methods have transitioned from using sparse point cloud to 4D radar tensors. However, the scarcity of publicly available 4D radar tensor datasets limits model generalization across diverse driving scenarios. Previous methods addressed this by synthesizing radar data, but the outputs did not fully exploit the spatial information characteristic of 4D radar. To overcome these limitations, we propose LiDAR-to-4D radar data synthesis (L2RDaS), a framework that synthesizes spatially informative 4D radar tensors from LiDAR data available in existing autonomous driving datasets. L2RDaS integrates a modified U-Net architecture to effectively capture spatial information and an object information supplement (OBIS) module to enhance reflection fidelity. This framework enables the synthesis of radar tensors across diverse driving scenarios without additional sensor deployment or data collection. L2RDaS improves model generalization by expanding real datasets with synthetic radar tensors, achieving an average increase of 4.25\% in and 2.87\% in across three detection models. Additionally, L2RDaS supports ground-truth augmentation (GT-Aug) by embedding annotated objects into LiDAR data and synthesizing them into radar tensors, resulting in further average increases of 3.75\% in and 4.03\% in . The implementation will be available at https://github.com/kaist-avelab/K-Radar.

Paper Structure

This paper contains 29 sections, 4 equations, 3 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Related work
  3. Radar Data Generation from Different Modalities
  4. Deep learning-based Synthesis
  5. Physics-based Simulation
  6. L2RDaS Base Model: Image Translation
  7. Method
  8. L2RDaS Generator Model
  9. Model Architecture
  10. Objective Functions
  11. Conditional Adversarial Loss
  12. Feature Matching Loss
  13. L1 Loss
  14. OBIS Module
  15. Adding Object Boundary Points
  16. ...and 14 more sections

Figures (3)

  • Figure 1: Application of the L2RDaS. L2RDaS is a framework designed to synthesize 4D radar tensors from LiDAR data. It can be applied to existing autonomous driving datasets that lack 4D radar tensors, enabling dataset expansion without additional sensor deployment or data collection. This allows models to achieve better generalization across diverse driving scenarios. Furthermore, L2RDaS supports ground-truth augmentation (GT-Aug) by first augmenting the LiDAR data and then synthesizing it into 4D radar tensors. The resulting tensors preserve realistic object and clutter measurements, allowing GT-Aug to be effectively applied to 4D radar tensor datasets as well.
  • Figure 2: Overall L2RDaS framework. The core components of the framework are the L2RDaS Generator and the OBIS module.
  • Figure 3: Qualitative results of the proposed L2RDaS framework. (a) Comparison with GT tensors on the K-Radar test set. (b) Results of applying L2RDaS to existing autonomous driving datasets that do not provide 4D radar tensors. (c) GT-Aug results; orange boxes indicate newly added objects. (d) Effect of the OBIS module in the ablation study.