Learning structural consistency and monocular priors for progressive depth completion

Depth completion fuses sparse depth measurements with RGB images to reconstruct dense metric depth maps. However, existing methods relying primarily on sparse annotations struggle to maintain globally consistent geometry. While recent monocular depth estimation models provide dense structural priors, their inherent scale ambiguity renders direct numerical alignment unreliable, risking the propagation of teacher bias. To overcome this, we reformulate depth completion as a supervision density augmentation problem and propose SCPNet. Central to this paradigm is a novel Order Transitive Ranking Loss that explicitly converts relative depth predictions into globally consistent transitive ordinal constraints, which serve as dense structural supervision complementary to sparse metric measurements. This paradigm fundamentally bypasses error-prone numerical fitting, enabling reliable sparse metric measurements to correct local inconsistencies. To faithfully accommodate these dense priors, SCPNet employs a dual-path structure-preserving encoder to decouple long-range geometric dependencies from boundary-sensitive representations, coupled with an anchor-guided multi-scale refinement strategy to ensure sparse observations remain geometrically dominant. Extensive experiments on KITTI, NYUv2, and VOID benchmarks demonstrate that SCPNet achieves state-of-the-art performance, exhibiting exceptional structural fidelity and robustness under highly sparse input conditions.