Demo 04 · Client-side

Object detection in the browser

Upload an image or point your webcam at something, and a YOLOX-Nano model finds the objects and draws the boxes — on your machine, in this tab. The readout breaks the run into preprocess, inference, and postprocess time, because in detection work the milliseconds are half the job. My internships at Mitsubishi Electric and DIATICS were exactly this kind of model, tuned for exactly these numbers.

Everything runs locally through onnxruntime-web (WebAssembly) — no API calls; images never leave the page. The model is YOLOX-Nano by Megvii (Apache-2.0, ~3.7 MB), trained on COCO: it knows those 80 everyday classes and nothing else, and at this size it will miss small or unusual objects — that tradeoff is the point of a nano model. First run downloads the model plus a ~13 MB wasm runtime; both are cached by the browser. Timings depend on your hardware.