GALMono: ambiguity-aware lightweight monocular depth estimation with graph convolution

Monocular depth estimation serves as a fundamental component in intelligent transportation systems and autonomous driving, supporting vehicles in perceiving and understanding the spatial structure of traffic scenes. High-capacity models yield excellent results but are computationally prohibitive for real-time deployment in embedded autonomous driving systems, whereas lightweight methods improve efficiency at the cost of reduced accuracy—particularly in object– background separation and ambiguous regions with limited texture or visibility. To address the above limitations, we design a compact self-supervised architecture that integrates graph convolution into the decoder for better structural reasoning and employs channel attention for refined feature extraction. Furthermore, a self-learning ambiguity boosting strategy is introduced to ensure consistent and robust depth estimation in challenging areas. Extensive experiments demonstrate that our method outperforms existing lightweight self-supervised baselines on the KITTI dataset, with a 5.49% improvement in the Sq Rel metric. Zero-shot evaluations on the Make3D dataset further highlight its strong cross-dataset generalization, accurately recovering distant and fine-grained structures. The method achieves both reliable performance and efficiency, delivering a viable option for on-the-fly monocular depth estimation within autonomous vehicle systems.