Abstract
Stereo endoscopy provides metric depth through binocular geometry and is an important sensing modality for robot-assisted minimally invasive surgery. However, surgical stereo matching remains challenging due to narrow baselines, textureless tissue, specular reflections, non-Lambertian appearance, and instrument occlusion. Meanwhile, stereo matching has rapidly advanced with monocular depth priors, large-scale pretraining, and foundation stereo models, yet their effectiveness in surgical scenes has not been systematically compared under a unified protocol. In this work, we introduce StereoSurg, a benchmark of recent stereo depth estimation methods, including IGEV++, MonSter, Mon- Ster++, S2M2, FoundationStereo, Fast-FoundationStereo, BridgeDepth, and StereoMamba. We use SCARED as the quantitative benchmark and evaluate disparity accuracy using EPE, D1, Bad-k metrics, and metricdepth error. We further compare computational cost in terms of model size, inference speed, and GPU memory, and qualitatively evaluate challenging in-vivo scenes from Hamlyn and StereoMIS. Under our unified protocol, FoundationStereo achieves the best EPE at 2.659 px and the lowest Bad1 and Bad3 rates, while BridgeDepth achieves the best D1 at 20.63%. StereoMamba reports a slightly lower EPE of 2.64 px and the best Bad2 at 41.49% under its SCARED fine-tuning protocol. Fast- FoundationStereo performs best on Dataset 8 with a mean depth error of 2.83 mm, whereas S2M2 performs best on Dataset 9 with 1.95 mm. Fast- FoundationStereo also achieves 28.02 FPS with 1.08 GB GPU memory. Qualitative comparisons reveal additional failure modes around specularities, weak texture, surgical instruments, and degraded regions. Overall, the benchmark highlights the trade-offs between accuracy, in-vivo robustness, and computational efficiency in surgical stereo estimation.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/CREATE_003.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to Open Review
Open Review Page: https://openreview.net/forum?id=DjcxQiY3Do
BibTex
@InProceedings{GuoJia_StereoSurg_MICCAISAT2026,
author = { Guo, Jiaxin AND Dong, Wenzhen AND Guan, Tongfan AND Wang, Ziyi AND Dou, Qi AND Liu, Yun-Hui},
title = { { StereoSurg: A Benchmark for Stereo Depth Estimation in Surgical Scenes } },
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}
}