Hybrid CNN–transformer for monocular depth estimation via adaptive region- aware embedding

Abstract Monocular depth estimation is a key component of autonomous driving, enabling 3D scene understanding from single images without relying on expensive sensors. Despite recent advances with hybrid CNN–Transformer architectures, existing approaches still struggle with semantic disparities between central and peripheral regions, spatially non-uniform traffic scenes, and effective cross-architecture feature fusion. To address these issues, we present DriveDepth, a hybrid framework designed specifically for depth estimation in complex driving environments. Its core component is an Adaptive Region-Aware Embedding (ARAE) module, which employs differentiable region masks to guide the model toward semantically critical areas of the scene, ensuring that key regions receive appropriate emphasis during feature extraction. In addition, DriveDepth incorporates a lightweight multi-scale attention module and a deformable feature enhancement module to further improve feature interaction and representation. On the Cityscapes dataset, DriveDepth achieves 4.571 RMSE and 90.6% δ < 1.25, reducing RMSE by 22% compared with the best existing method, and reaches 3.521 RMSE on KITTI, a 4.2% improvement over the best existing method. These results demonstrate that DriveDepth provides accurate and efficient depth estimation for complex real-world driving scenes.