Abstract
Accurate endoscopic 3D reconstruction is important for intraoperative scene understanding, navigation, and robotic assistance. Recent feed-forward visual geometry models generalize well in natural scenes, but their direct transfer to endoscopic videos is hindered by specular highlights, weak textures, tissue deformation, narrow fields of view, and imaging degradations. We formulate endoscopic 3D reconstruction as domain adaptation of a pre-trained visual geometry model and adapt VGGT through supervised depth-pose fine-tuning, perturbationenhanced training, and teacher-student geometric consistency. Experiments on multiple public endoscopic datasets show that endoscopic finetuning substantially improves camera pose estimation over zero-shot transfer, while perturbation augmentation and teacher-student consistency provide complementary robustness gains. These results suggest that general visual geometry priors can effectively support endoscopic reconstruction when combined with domain-specific adaptation.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/EndoLINA_012.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: https://papers.miccai.org/miccai-2026-sat/supp/EndoLINA_012_supp.pdf
Link to Open Review
Open Review Page: https://openreview.net/forum?id=cu3hdmtlhQ
BibTex
@InProceedings{LiuYan_Adapting_MICCAISAT2026,
author = { Liu, Yanzhe AND Yu, Yicheng AND Wu, Junyang AND Gu, Yun},
title = { { Adapting Generalizable Visual Geometry Models for Endoscopic 3D Reconstruction } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17275},
month = {pending},
page = {pending}
}
