OwlMask

Coverage

What to mask

Heads from body pose (OwlGuard method, covers turned-away heads) Faces (face detector, catches small faces) License plates Custom zones (drawn on the picture; fixed for every frame)

Wrong mask on a photo? Click it to exclude it. Too many false positives: raise the confidence sliders under Detection settings.

Mask style

Pixel blocks across each mask 8

Mask size ×1.0

Head size

Solid fill is the only style that cannot be reversed. Pixelation and blur hide identity from people, not from every recognition model.

Custom zones

Review: click a detection to exclude or restore it (photos, browser mode). Draw zone: drag on the picture to mask a fixed area such as a window, a screen or a photo frame; zones are sent to the server too.

Advanced settings

Coverage sets all of these at once; change one and the coverage shows "Custom".

Video

Mask hold 0.25 s Each mask stays on for this long before its first and after its last detection, so a head that is missed for a moment stays covered; the same person is also followed through brief misses. 0 turns the hold off.

Keep following a lost face or plate 2 s When the detector loses someone — a face turned into a pillow, a plate too far to read — the server keeps following that patch of the picture for up to this long, until it is gone or found again. 0 turns it off. (Server masking only.)

Detect every frame Every frame is the most accurate. Above 1, positions are interpolated between the analysed frames; faster, and still follows motion.

Detection

Person confidence 0.40

Face confidence 0.60

Plate confidence 0.30

Detection accuracy

Show detection boxes on the preview Show frames while a video is processed (a few per second; the saved video keeps every frame) Use the GPU (WebGPU) when the browser offers it (checked against the CPU result at start-up; takes effect after a reload)

Method

Heads: YOLO26-pose finds every person; the head centre is the mean of the nose, eye and ear keypoints (confidence ≥ 0.30). "Fitted" sizes the mask from the ear, eye and shoulder spacing; "OwlGuard 14%" uses the original square with a radius of 14% of the body length (min 18 px at 640). Both fall back to the top 22% of the body box when no facial keypoint is visible. Ported from owlsense/modules.py · blur_heads().

A YOLO26n COCO detector runs alongside: a pose "person" counts only when it also sees a person there (this drops the giraffes and ornaments the pose model mistakes for people), and plates are searched only inside the vehicles it finds. Faces: YuNet (OpenCV zoo, MIT), boxes grown to the whole head; a face without a person around it must be very confident or crowd-sized. Plates: YOLOv9-t 512 (open-image-models, MIT). Video is decoded frame by frame with WebCodecs, every frame is detected and masked, then re-encoded with the original audio copied over. Nothing is uploaded; saved photos carry no EXIF.

Drop photos or a video here JPEG, PNG, WebP · MP4, WebM, MOV. Several photos at once is fine. Files stay on this device.

OwlMask

OwlMask

Simple pricing

Sign in