Towards Continual Federated Learning of Monocular Depth for Autonomous Vehicles

Abstract

Recent investigations in computer vision for autonomous vehicles have focused on depth estimation from images, owing to its cost-efficiency and adaptability. Monocular depth estimation, using a single camera, is notable for its adaptability compared to binocular techniques that require two fixed cameras. Sophisticated methodologies employ self-supervised deep neural networks, while latest research proposes the use of federated learning to tackle crucial challenges for autonomous driving, such as data privacy, network usage, computation distribution, and connectivity robustness. Nevertheless, continual learning presents an additional challenge for both centralized and federated training, as updating models with novel datasets may induce the forgetting of previously acquired samples, thereby diminishing the model's versatility. Meanwhile, recent research indicates that experience replay strategies can be used to alleviate forgetting in autonomous driving use cases. We introduce ERFedSCDepth, an innovative approach amalgamating experience replay, federated learning, and deep self-supervision to allow the training of monocular depth estimators with high effectiveness and efficiency. Assessment using KITTI and DDAD datasets shows the efficacy of our approach, achieving enhanced continual learning performance over state-of-the-art baseline at standard depth metrics.