← News

ResearchSources: arXiv · 19 Nov 2026

SurgCBVR benchmark tests whether AI can retrieve surgical video by content

A new arXiv benchmark, SurgCBVR, targets video-to-video retrieval in surgery across eight public datasets, and its authors report that current models struggle badly.

In plain words

Hospitals record huge amounts of surgical video, but there has been no standard way to search these archives by content — for example, finding similar operations or complications. SurgCBVR is a published test set designed to measure how well AI models can retrieve relevant surgical clips. In the authors' evaluation, today's models perform poorly.

What was announced

SurgCBVR, described in an arXiv paper by its authors as the first benchmark for content-based video-to-video retrieval in surgery, combines eight public datasets spanning ophthalmic, laparoscopic, robotic-assisted, and colonoscopy domains. It covers six procedure types, 679 cases, and 76,489 clips totalling 106 hours. The benchmark has two tiers: Scene, which retrieves the same operative state across cases and sites, and Workflow, which retrieves and localizes phase transitions and deviations.

The authors evaluated six generic and surgical foundation models, two video retrieval methods, and contrastive heads trained on the benchmark, using within-dataset, cross-dataset, and pooled-gallery retrieval restricted to the same procedure type. They report that frozen representations barely beat a timing-only baseline, transition retrieval approaches chance, and trained heads lose up to half their accuracy when transferred across datasets — findings they say expose substantial limitations in current video representations.

What it does not show

  • The source is the paper's abstract only; full methods, per-model results, and peer-review status are not available in the supplied excerpt.
  • All performance findings are the authors' own reported results and have not been independently confirmed.

Sources

Source-support review completed 20 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 20 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.