Improving Domain Generalization via Bridging Convolutional Neural Networks and State-Space Models in Monocular Depth Estimation

Self-supervised monocular depth estimation (MDE) has emerged as a cost-effective alternative to traditional depth sensing approaches, which often require expensive equipment or complex configurations. However, these models often fail to maintain high performance on unseen domains, presenting a considerable issue for real-world applications, such as autonomous driving, where the ability to adapt consistently across various conditions is essential for ensuring safety. In this paper, we propose a novel lightweight network designed to separate domain-specific style and domain-invariant content within feature representation statistics, removing the style information that contributes to the domain shift. Furthermore, to address the convolutional neural network’s limited local scope, we inject state-space model blocks inside our encoder architecture in order to capture extensive contextual information, yet having a minimum number of trainable parameters. Experiments on popular outdoor depth estimation benchmarks, KITTI and Make3D, demonstrate superior performance to other recent networks with similar size.