Lab · Object detection

See what’s in the frame

A small in-browser demo that draws boxes and labels on an image or webcam view. Built so listeners of Field Notes with Kpo can try the idea behind YOLO-style detection — without sending your pixels to a server.

Runs on-device via TensorFlow.js COCO-SSD (Apache-2.0). Not Ultralytics YOLO weights — see licensing note below.

Podcast: 26W40-EP02 — episode coming soon

Try it

Upload a still photo, run Live webcam detection, or Capture one frame (lighter on phones). Nothing leaves this tab.

Loading model…

Samples

Curated cases — this list is meant to grow over time (assets/samples.json).

What this is

“YOLO” (You Only Look Once) is a family of real-time detectors that look at an image in one pass and propose boxes + labels. This page uses a related, browser-friendly model (COCO-SSD) so you can feel the idea: what it catches, what it misses, how lighting and angle matter.

It is a teaching sample, not a security product. Clear scenes work better than clutter. A pretty demo is not a production system.

Live keeps detecting every moment; Capture freezes one frame, then stops the camera so phones stay cooler.

Privacy

Nothing is uploaded from this page for detection. Images and webcam frames are processed locally in your browser. The model files load from public CDNs / model storage on first use (same as other TensorFlow.js demos). Stop the webcam when you are done; closing the tab ends the session.

Libraries & license

@tensorflow/tfjs@4.22.0 — jsDelivr CDN (Apache-2.0)
@tensorflow-models/coco-ssd@2.2.3 — jsDelivr CDN (Apache-2.0)
Default weights: SSD Lite MobileNet v2 via TensorFlow.js model hosting

Why not Ultralytics YOLO here? Those releases are AGPL-3.0 (or Enterprise). For a closed personal site without AGPL’ing the whole project, a permissive stack is safer. Details: licensing notes.

Field Notes

26W40-EP02 — YOLO, Plainly — and a Demo You Can Try — episode coming soon on Spotify (Field Notes with Kpo).

← Home