Nov 2026· Proceedings of the 35th ACM International Conference on Information and Knowledge Management· 0 citations· 27 references
TL;DR
Analysis of routing utilization, top-k selection, trainable parameters, and inference latency demonstrate that IC-DVLoRA provides an effective and efficient adaptation strategy for endoscopic depth foundation models.
Abstract
Transferring depth foundation models to endoscopic videos is promising for surgical geometric perception, but remains challenging under patient-level visual heterogeneity. Natural-scene depth models provide strong geometric priors, yet endoscopic imagery differs substantially in specular reflection, weak texture, tool occlusion, close-range anatomy, and acquisition conditions. Existing parameter-efficient adapters, including DV-LoRA-based methods, are typically static after training and apply the same adaptation to all frames, limiting their ability to handle input-dependent appearance shifts. We propose IC-DVLoRA, an input-conditioned dynamic adaptation framework for endoscopic depth foundation models. For each frame, a lightweight condition encoder extracts a compact token that captures appearance, texture, gradient, and structural cues. This token dynamically routes multiple DV-LoRA experts in a frozen transformer backbone and modulates a DPT-style decoder through feature-wise transformations, enabling sample-specific low-rank adaptation while preserving pretrained geometric knowledge. We train and validate IC-DVLoRA on SCARED and evaluate its zero-shot generalization on Hamlyn and C3VD without target-domain fine-tuning. Experiments show improved overall performance over static adaptation baselines and competitive results against recent endoscopic depth estimation methods. Further analyses of routing utilization, top-k selection, trainable parameters, and inference latency demonstrate that IC-DVLoRA provides an effective and efficient adaptation strategy for endoscopic depth foundation models.
Through the first systematic study of TTA for dense depth regression (eight baselines), it is shown 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.
Sheng-Yao Wang, Jun-Pei Xue, Xin Zhong et al.· IEEE Transactions on Medical...· 1 citation
The proposed CRFDepth framework effectively addresses key challenges in endoscopic monocular depth estimation, achieving superior accuracy and robustness.
Yang-Yi Deng, Can Yan, Nai-Jung Cheng· Scientific Reports· 0 citations
Dynamic Gaussian reconstruction requires compact scene features to support coordinated changes in geometry and appearance. A factorized grid provides efficient local queries, but accurate interpolation alone does not ensure that the resulting dynamic attributes remain compatible across space and time. We introduce DimF...
The key insight is that DA3's cross-view features, though trained only for depth consistency, implicitly encode motion-discriminative signals when combined with best-match feature search across frames, which could enable much denser temporal sampling that prior methods cannot support.
Xin-Hao Xiang, Wei-Yang Li, Zhi-Jie Zheng et al.· 0 citations
Robust 4D reconstruction of deformable endoscopic scenes from monocular videos remains challenging due to non-rigid tissue deformation, illumination changes, occlusions, and unknown camera trajectories. Reliable dynamic reconstruction of endoscopies could significantly enhance visualization, diagnostic accuracy and int...
Laura Salort-Benejam, S. Giannarou, A. Agudo· 0 citations