SPIdepth: A Novel Approach to Enhancing Monocular Depth Estimation with Strengthened Pose Networks

Self-supervised monocular depth estimation is a rapidly advancing field with significant implications for autonomous driving and robotics. However, many existing approaches fail to fully leverage the potential of enhancing pose information, which is crucial for accurate depth estimation. In this paper, we introduce SPIdepth, a novel method that enhances the pose network to improve depth estimation performance. By building on the Self Query Layer (SQL) framework, SPIdepth places a strong emphasis on capturing fine-grained scene structures through more effective use of pose information, resulting in notable improvements in scene understanding. Our experimental evaluation on the KITTI benchmark dataset demonstrates that SPIdepth achieves state-of-the-art performance, outperforming previous methods by substantial margins. Specifically, SPIdepth sets new records on the self-supervised KITTI benchmark, as well as sets new state-of-the-art on KITTI with achieving the lowest AbsRel (0.029), SqRel (0.069), and RMSE (1.394) metrics. Notably, SPIdepth achieves these results using only a single image for inference, outperforming even those methods that rely on video sequences, showcasing its efficiency and effectiveness in practical applications. Although primarily designed for autonomous driving, SPIdepth’s robust depth estimation and scene understanding capabilities also have significant potential for Earth observation (EO) applications. This makes it particularly valuable for geospatial analysis, disaster management, and recovery efforts, such as those required for Ukraine's post-conflict reconstruction, where precise environmental monitoring and informed decision-making are critical. The code and pre-trained models are publicly available at https://github.com/Lavreniuk/SPIdepth.