# FaceOff: Anonymizing Videos in the Operating Rooms

arXiv: https://arxiv.org/abs/1808.04440 · HTML: https://mlsanity.com/paper/1808.04440/

## Key facts

- Cited by 5 papers in the map (data through 14 Oct 2026).
- First posted on arXiv 6 Aug 2018.

## Built on by (most cited first, 5 of 5)

- [DisguisOR: Holistic Face Anonymization for the Operating Room](https://mlsanity.com/paper/2307.14241/) (2023)
- [Beyond Role-Based Surgical Domain Modeling: Generalizable Re-Identification in the Operating Room](https://mlsanity.com/paper/2503.13028/) (2025)
- [SafeTriage: Facial Video De-identification for Privacy-Preserving Stroke Triage](https://mlsanity.com/paper/2506.16578/) (2025)
- [Automatic Detection of Out-of-body Frames in Surgical Videos for Privacy Protection Using Self-supervised Learning and Minimal Labels](https://mlsanity.com/paper/2303.18106/) (2023)
- [Self-Supervised Uncalibrated Multi-View Video Anonymization in the Operating Room](https://mlsanity.com/paper/2602.02850/) (2026)

## Abstract (arXiv)

Video capture in the surgical operating room (OR) is increasingly possible and has potential for use with computer assisted interventions (CAI), surgical data science and within smart OR integration. Captured video innately carries sensitive information that should not be completely visible in order to preserve the patient's and the clinical teams' identities. When surgical video streams are stored on a server, the videos must be anonymized prior to storage if taken outside of the hospital. In this article, we describe how a deep learning model, Faster R-CNN, can be used for this purpose and help to anonymize video data captured in the OR. The model detects and blurs faces in an effort to preserve anonymity. After testing an existing face detection trained model, a new dataset tailored to the surgical environment, with faces obstructed by surgical masks and caps, was collected for fine-tuning to achieve higher face-detection rates in the OR. We also propose a temporal regularisation kernel to improve recall rates. The fine-tuned model achieves a face detection recall of 88.05 % and 93.45 % before and after applying temporal-smoothing respectively.