Boosting self-supervised multi-frame depth estimation with hybrid geometric-semantic constraints

Self‑supervised multi‑frame monocular depth estimation leverages semantic appearance and geometric matching information to improve depth prediction through temporal cues, yet it continues to face challenges posed by semantic and geometric inconsistencies in low‑texture, occluded, and dynamic regions. While existing methods often rely on spatial context to address such inconsistencies, they typically overlook explicit geometric priors. To overcome this limitation, we introduce an adaptive hybrid geometric and context‑aware attention module, which refines corrupted cost volumes by aggregating local geometric and semantic information and propagating consistent neighboring costs guided by surface normals and epipolar geometry. Complementing this design, we propose an iterative multi‑scale hybrid decoder that effectively fuses both multi‑frame and single‑frame cues encoded in the cost volumes. Extensive experiments across diverse real‑world scenarios, including outdoor driving datasets (KITTI, Cityscapes), indoor environments (NYUv2, ScanNet), and challenging adverse‑weather conditions (DENSE), validate the robustness, accuracy, and strong generalization capability of the proposed method.