Validation notes / 2026-09-04
Model review finds disagreements; it cannot replace human ground truth. This page separates prepared regression assets from real-world datasets still requiring permission, download or annotation. External links are for reference and never automatically send your video anywhere.
Try a cinema close-up, two people turning quickly, people approaching the camera, and a blurred indoor clip. Sources are Blender Foundation films under CC BY 3.0 and Intel videos under CC BY 4.0. Audio has been removed. Artificial speed-up and blur are labeled in the sample picker.
These samples have no frame-by-frame human ground truth. Use them to explore the tool and look for missed faces, not as accuracy benchmarks.
Authors, sources and modifications ↗A NASA astronaut photo provided by scikit-image generates translation, small-face, partial-occlusion, crossing, scene-cut and no-face clips. Known positions in each frame help check coordinates, timestamps, masks and export. These are synthetic, not evidence of real-world accuracy.
Developers run npm run test:integration. Synthetic assets and full regression results stay in the developer's local work/validation directory and are not published with the site.
Useful for speech, head turns, hand occlusion, indoor lighting and audio/video synchronization. The official corpus provides about 100 hours of meetings under CC BY 4.0. Older DivX AVI downloads need local conversion to a browser-supported format.
Status: source and license checked; full corpus not downloaded. Short clips and complete frame-by-frame face boxes are still needed. Speech and head-motion annotations are not complete face-detection ground truth.
Official download ↗ · License ↗Useful for distance, scale, yaw, occlusion and repeated entry/exit. The source provides original frame sequences and eye-position annotations. Test full frames, not cropped faces.
Status: checked, not downloaded. Licensing is limited to noncommercial research and personal experiments; confirm permission before commercial product evaluation. Eye annotations are not full face boxes.
Source, license and download ↗Useful for crossing people, small faces, long occlusions and track-link errors. The authors provide scene reconstruction steps and per-frame box/track annotations. Some videos depend on YouTube or ChokePoint; a public repository does not imply commercial video rights.
Status: checked, not downloaded. Annotations include detector outputs and flags. Check completeness; do not assume independent ground truth covering every face.
Author dataset and annotation notes ↗Useful for individual pose, quality and clip variation. Its primary task is face verification, not full-frame missed-face evaluation. Aligned face crops cannot establish general video detection recall.
Status: checked, not downloaded. Official access requires a form; no personal details have been submitted. Video rights need separate verification.
Official database ↗Without human validation on natural video datasets, we do not claim zero missed faces, 99% accuracy or production readiness.