Leveraging Bayesian Optimization to Enhance Self-Supervised Federated Learning of Monocular Depth Estimators

Abstract

Current investigations in computer vision on intelligent transportation systems have notably centered on depth estimation from images, primarily because of its cost-efficiency and diverse range of applications. Monocular methods for estimating depth have garnered significant interest due to their dependence on a single camera, which provides considerable versatility compared to binocular techniques necessitating two fixed cameras. Despite leveraging sophisticated self-supervised deep neural network learning techniques with surrogate tasks such as pose estimation and semantic segmentation, essential requirements for the practical deployment of autonomous vehicles are often neglected. These entail safeguarding data privacy, minimizing network usage, distributing computational expenses, and fortifying resilience against connectivity challenges. Recent investigations underscore the efficacy of combining federated learning with Bayesian optimization to meet these criteria without compromising the model’s effectiveness. Therefore, we present BOFedSCDepth, an innovative approach that combines Bayesian optimization, federated learning, and deep self-supervision to train monocular depth estimators with improved effectiveness and efficiency compared to the current state-of-the-art method in self-supervised federated learning. Evaluation experiments conducted on the KITTI dataset show- case the superiority of our method, achieving a communication cost reduction of up to 33.3% and linear computational costs at the central server with no extra burden on autonomous vehicles.