Can Image Style Transfer Save Automotive Radar?

Abstract

Compared to RGB camera and Lidar, single chip automotive radar is a promising alternative sensor with robustness to adverse weathers. But the sparseness of radar output drastically hinders its usefulness for autonomous driving tasks. Up-sampling via image style transfer could be a cure for a sparse measurement. However, it remains unknown whether style transfer can be an effective solution to automotive radar which features different and unique sparse and noisy issues. In this paper, we evaluate a variety of predominant image style transfer methods for a typical ego-vehicle pose estimation task on the public nuScenes dataset, and find that though image style transfer methods can improve the visual quality of automotive radar measurements, they can hardly contribute to the utility of radar for downstream tasks.