Self-supervised learning of monocular depth estimators in autonomous vehicles with federated learning

In recent years, artificial intelligence has been applied to improve intelligent transportation systems, with a special focus on developing autonomous vehicles. Monocular depth estimation is a crucial task in autonomous driving, offering a cost-effective alternative to binocular methods. While existing approaches for this task leverage self-supervised learning of deep neural networks, they have not taken into account key requirements for their use in autonomous vehicles, such as the need to preserve the privacy of samples collected by multiple vehicles, reduce network consumption during training, and optimize the computation cost distribution. In recent studies, the use of federated learning has shown notable benefits when addressing these requirements. Thus, we propose a novel method combining federated learning and deep self-supervision to enable training of monocular depth estimators with comparable efficacy and superior efficiency to current methods. Our evaluation experiments, using two public benchmarks of images captured by moving vehicles, show that our proposed method achieves near state-of-the-art performance, with test losses of 0.115 and 0.169 on each dataset requiring, on average, only 1k to 1.25k training steps and about 65.1% to 87.4% less data transfer per vehicle in each round than our baseline.