Self-supervised Depth Estimation for Monocular Endoscopy Using Confidence-Rectified Distillation and Semantic...

Monocular depth estimation (MDE) is a core capability for 3D reconstruction and intraoperative augmented reality in endoscopy. Recent self-supervised methods improve performance by leveraging reprojection consistency, temporal constraints, and stronger representations. However, endoscopic images exhibit specular highlights, low texture, and pronounced non-Lambertian effects. These factors introduce noise, degrading depth accuracy in endoscopic settings. In this work, we further introduce two complementary innovations, confidence-rectified distillation and semantic distribution alignment (SDA), which jointly improve accuracy and strengthen cross-domain generalization. Specifically, a frozen teacher produces multi-resolution pseudo-labels, and we align them globally in the log domain, converting multiplicative scale gaps into an additive bias and improving robustness to specular outliers. We further introduce a Multi-Resolution Prediction Consistency Filter (MRPCF) that estimates per-pixel confidence from cross-scale log-depth variance, thereby realizing confidence-rectified distillation and applying soft weighting to suppress unreliable supervision. In parallel, we freeze DINOv3 as a semantic prior and equip the semantic teacher with a lightweight MLP projection head that maps semantic features to the student's geometric feature space for spatial alignment. The student is optimized with a temperature-scaled Kullback-Leibler (KL) divergence to match semantic distributions, which improves boundary fidelity and preserves thin structures. On the official SCARED split, our approach surpasses existing self-supervised methods. Zero-shot evaluations on the Hamlyn Heart and SERV-CT datasets without fine-tuning also deliver consistent gains, demonstrating strong cross-dataset generalization.