Spatio-temporal consistency for semi-supervised learning using 3D radar cubes

Radar has been employed as a key component of perception modules for more than two decades. However, radar image labelling requires expert knowledge. At the same time, it is much more time-consuming than for general RGB images, which impedes further developments of radar. In order to alleviate the high-cost annotation problem in radar datasets, we present a novel, semi-supervised deep learning method based on the spatio-temporal consistency. This way we explore the potential of unlabeled radar frames to enhance performance. We utilize the consecutive radar frames from different timeline directions to encourage the model to learn the target motion. Moreover, the proposed self-weighted mechanism avoids over-fitting on certain predominant targets, by exploiting the supervised classification loss dynamically. We evaluate the proposed method on semantic segmentation and Vulnerable Road Users (VRUs) detection problems. The quantitative results compare favourably to the state-of-the-art and demonstrate the effectiveness of the proposed concepts. The ablation studies also show the effectiveness of the proposed components.