SAUC-Depth: Specularity-Aware Uncertainty-Calibrated Monocular Depth Estimation for Endoscopic Scenes
Abstract
Accurate monocular depth estimation in endoscopic scenes is challenging because of specular reflections, weak texture, illumination variation, and uncertainty in image-derived depth cues. This study presents SAUC-Depth, a specularity-aware and uncertainty-calibrated framework that integrates uncertainty modelling, reliability-weighted supervision, boundary preservation, and cross-resolution consistency for monocular endoscopic depth estimation. Experiments are conducted using publicly available rectified endoscopic sequences with sequence-independent training, validation, and testing. Performance is assessed using standard depth-estimation metrics and uncertainty reliability analysis. The SAUC-Depth method reports an AbsRel of 0.110, RMSE of 8.637 mm, MAE of 6.347 mm, RMSElog of 0.148, and δ1 of 0.880. The proposed framework focuses on improving robustness in visually ambiguous regions while providing pixel-wise reliability information for downstream surgical vision applications.