How it works, and what SeeDeep was
Everything the page does, with its sources, and the honest version of the history.
The detector
MediaPipe Tasks Vision, Google's library for running vision models in the browser, under the Apache 2.0 licence, version 0.10.21, with the EfficientDet-Lite0 model (float16, 7.3 MB) trained on the COCO dataset's 80 classes. The runtime is WebAssembly; on browsers that allow it the model runs on the graphics processor, otherwise on the processor. Detections below a score of 0.4 are dropped; at most 25 are kept per frame. Sources: the MediaPipe object detector page, the licence, and the files themselves at /vendor/ and /models/.
Tracking and counting
Our own code, about 150 lines, in /js/seedeep.js. A detection in one frame is matched to a track from the last frames when the boxes overlap by more than 30 percent and the class is the same; a track is counted as seen after three matched frames; it is dropped once it has been missed in more than twelve frames. The line starts at 62 percent of the frame's height and can be dragged by its ends to any place and angle; a crossing is counted when a track's centre passes from one side of the segment to the other, within its ends. The zone is a four-cornered polygon; a track is inside when its centre is, an entry is counted once per track, and a stay is the time from entry to exit. The line and zone are kept in your browser's own storage, for you only. Counts are honest about their nature: two people who swap places can be counted as one, and someone who leaves and comes back is counted twice.
Measured
On a MacBook Pro (Apple M1 Pro), Google Chrome 155, 21 October 2026: 36 frames a second, 27 ms a detection (GPU delegate). The page shows your own device's numbers live, and the "Copy these numbers" button puts them on your clipboard with the date and your browser's name. We have not measured the model's accuracy ourselves; MediaPipe publishes its figures on the page linked above.
What stays on your device
All of it. The page's content rule forbids any request to another host except Cloudflare's visit counter and the link counter at sevenskyconsulting.com, and there is no server of ours behind it: the site is static files. Cloudflare, which serves the pages, counts page visits with its Web Analytics, which sets no cookie: page visits only, never camera frames or files. Opening your camera asks the browser's permission in the usual way; the stream is drawn to the page and nowhere else. A video file you open is read by the page through a local address the browser makes for it.
What SeeDeep was
In 2025 SeeDeep ran as a small service: a Python web service on one rented server with five detectors (face masks, caps, protective equipment, weapons, and the 80 everyday classes), and a marketing site that offered a hair colour changer, a vehicle counter and smart CCTV detection, with posts whose cost figures had no published source. The server stopped answering in August 2026 and the contract ended. The site as it stood on 25 February 2026 was copied from the Internet Archive and republished in October 2026 as a record, its unsourced figures removed. This version, built 21 October 2026, is the vehicle counter and the detection, running in front of you; the hair colour changer is the idea behind makeupme.ai.
Credits
- Sample clip: a diagonal crosswalk at Yonge and Dundas, Toronto, by Raysonho, Wikimedia Commons, CC0, cut to 24 seconds at 960 by 540.
- Typeface: Montserrat, by Julieta Ulanovsky and others, SIL Open Font Licence.
- Detector and runtime: MediaPipe, Google, Apache 2.0.
Contact: moe@sevenskyconsulting.com.