SurgDepth: Test-Time Adaptation of Depth Foundation Models for Surgical Scene Understanding.

Abstract

Accurate depth estimation from endoscopic images is fundamental to surgical navigation, augmented reality, and robotic assistance, yet depth foundation models trained on natural images degrade severely on surgical endoscopy, and existing adaptation methods require surgical training data for each new camera and clinical environment. We propose SurgDepth, a test-time adaptation (TTA) framework that adapts pre-trained depth foundation models to surgical endoscopy without any labeled surgical data. SurgDepth combines two self-supervised signals-stereo photometric consistency for rectified stereo pairs, and flip equivariance requiring only single images-with selective decoder adaptation and two tuning-free reliability mechanisms: a progress-aware anchor regularization and a structural-trust safeguard for poorly-illuminated sequences. Through the first systematic study of TTA for dense depth regression (eight baselines), we show that existing TTA methods consistently fail in this setting: entropy-based methods degrade regardless of the number of adapted parameters, and temporal-photometric adaptation diverges on non-rigid tissue. SurgDepth reduces AbsRel by up to 63% with stereo pairs and 42% without, improves four of five cross-domain evaluation settings (three of four unseen datasets; Hamlyn is evaluated under two protocols), generalizes across seven foundation models, and attains the lowest AbsRel (0.128) on the complete standard Hamlyn benchmark among all reported methods. All results evaluate relative (median-scaled) depth; 20-step convergence (0.4 s) makes SurgDepth practical for relative-depth applications such as augmented-reality overlay and reconstruction initialization.