A hands-on comparison of three face detectors found that YuNet caught every face in the author’s test images and short video, while another detector missed every small face in some wide shots. That matters for anyone using automatic face blurring: the blur only protects people the software actually finds.
Watch Desk analysis
What happened
Jacob Yang tested OpenCV’s Haar cascade, YuNet and ResNet-10 SSD on generated portraits, composite group scenes and a three-second video. The tests ran on CPU with fixed detection thresholds. Yang says the images were constructed with known face positions, not candid photographs, and the results describe these fixtures rather than every real-world setting.
In the still-image tests, YuNet detected all the ground-truth faces. Haar missed profile faces in two scenes and produced the same false positive on a patch of foliage across three scenes. ResNet-10 SSD detected none of the faces sized 28 to 36 pixels in one wide scene, and found only one of four faces in a group scene.
The video test had three faces in each of 90 frames. YuNet detected all 270 face instances. Haar missed 11, all involving a moving face. ResNet-10 SSD missed the moving face in every frame, leaving at least one face undetected each time. Yang reports that a five-frame hold helped Haar slightly but did not fix ResNet-10 SSD’s recurring miss.
Why it matters
A missed detection is not a cosmetic glitch in a face-blurring workflow. It leaves the face untouched. The comparison offers a practical warning for developers choosing a detector: performance on a large, clear portrait does not guarantee coverage in a wide shot, and a detector that works on one view may falter when a face moves or turns.
The numbers are useful, but their boundary matters. Yang tested a small set of generated and composited scenes, not a broad collection of real photographs or video. He recommends starting with YuNet for this fixture set, while noting that it also incorrectly detected an ear as a face in one image. No detector earns a universal crown from three seconds of video, even if one has made a very strong opening case.
Our read
The best takeaway is less “use this detector everywhere” and more “measure the failure that would matter in your application”. For privacy tools, recall is especially important: one sharp face can undo a lot of careful blur. Test small faces, profiles, movement and empty scenes, and inspect false positives as well as misses.
Yang’s comparison is an attributed community article on Hugging Face. Its detailed method and openly described limitations make it a useful starting point, not a substitute for testing on the footage a real system will handle.
What to watch
- Whether future tests include a larger and more varied set of real-world images and video.
- How detectors balance missed faces against false positives in practical blurring workflows.
- Whether developers publish repeatable benchmarks across different resolutions and operating conditions.
Discussion spark: For automatic face blurring, should developers prioritise catching every face even if that means more false positives, or is there a point where over-blurring makes the tool less useful?
Sources and evidence
- Face Blur in Python: What Three Detectors Miss on Profile, Small, and Group Faces (15 November 2026, 04:56 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.