Solutions
Lab YOLO update — searchable frames & a car-park Web App
Field Notes, continuing the YOLO thread.
Last episode I sketched a small idea: detect → describe → act, then jump to a timestamp instead of scrubbing video. This week is an update — with something you can try yourself.
I built a browser lab at https://kpoleung.com/lab/yolo/. It runs in the tab. No upload. No fee. Webcam or a local photo. Under the hood it’s TensorFlow.js with COCO-SSD (Apache-2.0) — YOLO-style, not Ultralytics on a public personal page, because AGPL gets awkward there. The scene line under the picture is built from the boxes the model already found: labels, counts, positions. There’s a Capture button when the camera is live. Honest limit: the demo only knows the usual everyday object set (about eighty COCO classes). Your own objects, uniforms, or actions need custom training — Colab and something like MediaPipe Model Maker fit that path, with a licence and export story that still aims at browser or phone.
Try the lab: https://kpoleung.com/lab/yolo/ — runs in your browser; nothing is uploaded for detection.
The bigger pattern, updated from last time: an agent watches an RTSP stream, grabs frames on an interval, writes text metadata with a timestamp, and later you search the text instead of watching hours. When a line hits, jump to that time in the NVR or RTSP recording — if playback-by-time is supported. Cameras are sensitive; keep retention and search access intentional.
Part B is new. Small on-device apps handle simple jobs; complex ones still need professional systems. Car-park patrols burn time copying plates and bay numbers. A phone Web App that detects, OCRs (plain COCO-SSD won’t read plates), logs plate + bay + photo, and works offline in a basement, can cut the copying. A guard confirms each read. Plates can be personal data — keep records local or controlled. I’ve built a sample, HK Plate Scanner, that reads plates on the phone: try it here (the first load downloads the model, then it’s cached).
Not a product pitch. A field note. The episode is voiced by two AI hosts — Holiday and Feiertag — on Field Notes with Kpo. Poke the lab if you’re curious.