Privacy-Preserving Training of Monocular Depth Estimators via Self-Supervised Federated Learning

Abstract

Monocular depth estimation is gaining attention in computer vision for autonomous driving due to its cost-effectiveness and versatility. Recent works leveraged self-supervised learning to tackle this problem. However, they overlook crucial challenges for real-world deployment, such as data privacy, network consumption, computational cost, and unstable connectivity. Recent studies have shown the potential benefits of federated learning in addressing these challenges. Thus, we introduce FedSCDepth, a novel method that combines federated learning and deep self-supervision to efficiently learn monocular depth estimators with performance comparable to the state-of-the-art. Our evaluation using the KITTI dataset's Eigen's Split reveals that FedSCDepth achieves high efficacy, with a test loss below 0.13. It also demonstrates impressive efficiency, requiring only an average of 1.5k training steps and up to 0.415 GB of weight data transfer per autonomous vehicle in each round.