Survey + benchmark Preprint · under review · 2026

Monocular vSLAM for Colonoscopy

Recent advances, benchmarking, and future perspectives

Duc Manh Do^1,3, Van Bang Nguyen¹, Thanh Nguyen Canh^1,2, Mai Thanh Thai^3,4,*, Xiem HoangVan^1,*

¹VNU University of Engineering and Technology ²Japan Advanced Institute of Science and Technology ³College of Engineering and Computer Science, VinUniversity ⁴VinUni Center for AI Research ^*Corresponding authors

Paper list See the benchmark Explore the systems

20: systems reviewed
13: methods re-run on one benchmark
4: colon segments (C3VDv1)
5: datasets compared
6: open questions

01 · Overview

A 3D map of the colon from ordinary colonoscopy video

Colonoscopy is the gold standard for colorectal cancer screening, but its narrow field of view limits spatial awareness, coverage assessment and lesion localization. Monocular visual SLAM can estimate the endoscope's motion and reconstruct the colon in 3D from standard video, with no extra hardware.

The colon is a hard place for SLAM: weak and repetitive texture, specular reflections, changing illumination and tissue deformation. This survey traces methods from classical feature-based and direct SLAM to learning-based, neural-rendering and deformation-aware systems, and adds the first unified benchmark of monocular vSLAM for colonoscopy: 13 methods re-run under one protocol on C3VDv1.

Overview of colonoscopic vSLAM: key challenges (weak texture, specular reflection, illumination change, tissue deformation), SLAM localization and mapping from colonoscopic video, and clinical applications (navigation assistance, lesion measurement, lesion localization, missing region detection). — Fig. 1: key challenges, workflow and clinical applications. Some non-data illustrative elements were made with AI assistance and edited by the authors; no experimental or patient data were generated.

Systematic review

Classical geometric, learning-enhanced and neural scene representations, and how each copes with colonoscopy's challenges.

Unified benchmark

Representative methods re-implemented and evaluated on the same data, comparing accuracy, reconstructions and speed.

Clinical readiness

Dataset realism, evaluation protocols and workflow integration on the road to clinical deployment.

02 · Foundations

Two classic pipelines

Feature-based SLAM minimizes reprojection error over matched keypoints; direct SLAM minimizes photometric error over image regions. Learning-based methods replace or strengthen parts of both.

Feature-based monocular vSLAM pipeline: RGB image, feature extraction and tracking, geometry constraint, keyframe decision, optimization, camera poses and 3D map. — Fig. 6: feature-based monocular vSLAM.

Direct monocular vSLAM pipeline: RGB image, image pyramid, keyframe decision, photometric constraint, optimization, camera poses and 3D map. — Fig. 7: direct monocular vSLAM.

03 · Evolution

From the first endoscope to neural colon maps

Historical milestones: 1806 Bozzini's Lichtleiter, 1983 first video endoscopy system, 2011 direct SLAM, 2015 feature-based SLAM, 2017 deep-learning-based SLAM, 2023 neural rendering for colonoscopy, 2026 this survey. — Fig. 4: historical milestones. Some non-data illustrative elements were made with AI assistance and edited by the authors.

The 20 systems of Table 5 by year and family. Hover or tap for the reference; click to open it in the explorer.

Bar chart of annual publications on 3D reconstruction and on SLAM for colonoscopy from 2002 to 2025, both rising sharply after 2020. — Fig. 2: annual publications on 3D reconstruction and SLAM for colonoscopy (keyword search on Elsevier with manual screening).

04 · Explorer

Explore the systems

Monocular vSLAM systems applied to, or applicable to, colonoscopy (Table 5). Press / to search; click a row for its reference.

Search systems Sort

System	Year	Family	Domain	Map	Deformable	Needs depth	Real-time	Code

05 · Unified benchmark

13 methods, one protocol, four colon segments

Positional accuracy (ATE and RPE, mm) and speed on C3VDv1 (Table 8). Choose a segment and a metric.

Reconstructions on the four C3VDv1 segments by ORB-SLAM3 (sparse), Endo-Depth-and-Motion, BodySLAM, EndoGSLAM, Endo-2DTAM and DROID-SLAM (semi-dense). — Fig. 9: reconstructions on C3VDv1. ORB-SLAM3 gives sparse points, DROID-SLAM semi-dense points; depth-prior and Gaussian-splatting methods give dense surfaces.

Lessons from the benchmark

Lesson 1

Dense and neural methods are sequence-dependent

EndoGSLAM and Endo-2DTAM are strong on sigmoid_t1_a but degrade on desc_t4_a, and their maps become unstable when localization fails.

Lesson 2

Depth helps globally, not always locally

In feature pipelines depth improves global alignment but not local motion; it must be coupled into tracking, and predicted depth needs explicit scale alignment.

Lesson 3

Datasets are short and slow

Current sequences cover short colon segments with slow motion, so they cannot test robustness over a full examination, motion blur or feature loss.

06 · Datasets

What data is available?

Table 4 of the paper. Non-commercial = open for non-commercial use.
Dataset	Type	Deformable	Year	Availability	Access	Depth GT	Trajectory GT	3D map GT

Evaluated C3VDv1 segments (Table 6)

Segment	Frames	Trajectory length (mm)

07 · Open questions

What the field still has to answer

08 · Cite

Citation

@misc{do2026monocular,
  title  = {Monocular vSLAM for Colonoscopy: Recent Advances, Benchmarking, and Future Perspectives},
  author = {Do, Duc Manh and Nguyen, Van Bang and Nguyen Canh, Thanh and Thai, Mai Thanh and HoangVan, Xiem},
  note   = {Preprint, under review},
  year   = {2026}
}