L2R GAN: LiDAR-to-Radar Translation

Leichen Wang, Bastian Goldluecke, Carsten Anklam; Proceedings of the Asian Conference on Computer Vision (ACCV), 2020

The lack of annotated public radar datasets causes difficulties for research in environmental perception from radar observations. In this paper, we propose a novel neural network based framework which we call L2R GAN to generate the radar spectrum of natural scenes from a given LiDAR point cloud.We adapt ideas from existing image-to-image translation GAN frameworks, which we investigate as a baseline for translating radar spectra image from a given LiDAR bird's eye view (BEV). However, for our application, we identify several shortcomings of existing approaches. As a remedy, we learn radar data generation with an occupancy-grid-mask as a guidance, and further design a set of local region generators and discriminator networks. This allows our L2R GAN to combine the advantages of global image features and local region detail, and not only learn the cross-modal relations between LiDAR and radar in large scale, but also refine details in small scale. Qualitative and quantitative comparison show that L2R GAN outperforms previous GAN architectures with respect to details by a large margin. A L2R-GAN-based GUI also allows users to define and generate radar data of special emergency scenarios to test corresponding ADAS applications such as Pedestrian Collision Warning (PCW).