LiDAR-EDIT: Lidar Data Generation by Editing the Object Layouts in Real-World Scenes
Shing-Hei Ho, Bao Thach, Minghan Zhu
Accepted to ICRA 2025
Embedded Files
Abstract
We present LiDAR-EDIT, a novel paradigm for generating synthetic LiDAR data for autonomous driving. Our framework edits real-world LiDAR scans by introducing new object layouts while preserving the realism of the background environment. Our experiments show that LiDAR-EDIT benefits the development and testing of autonomous driving system.
Advantages of LiDAR-EDIT:
Faithfulness: Generated data is only partial modification of real LiDAR scans, preserving characteristics of the original data
Controllability: Full control over object layout, including the number, types and poses of objects
Scalability: Generative instead of relying on physics engine, prebuilt 3D assets, or reconstruction of the whole scene
Method
Example LiDAR Editing
Original LiDAR scene:
Car removal and background inpainting:
Car insertion
Resampling with spherical voxelization
More Examples
Original LiDAR scenes:
Edited LiDAR scenes:
Page updated
Google Sites
Report abuse