A real detection model — not a video, not a mockup. Load an image or open your camera and watch it label objects frame by frame, with the actual inference latency reported next to it.
Pick an image, start your camera, or try the sample below.
The dial every real deployment argues over. Lower it to catch more objects and accept more false positives; raise it to keep only confident calls and miss the marginal ones. There is no universally correct value — it depends entirely on whether a miss or a false alarm costs you more.
This runs COCO-SSD (MobileNet v2), a general-purpose detector trained on 80 everyday object classes. It is deliberately not a custom model — the point is to show a working pipeline in a hostile environment, not to claim a benchmark.
Running inference in a browser tab means no GPU server, no batching, and whatever device the visitor happens to hold. Production work usually has more headroom than this — which is rather the point of showing it here.
It is measured around the actual forward pass on your hardware, not a stored figure. It will differ between your laptop and your phone, and that variance is exactly what capacity planning has to account for.
A production system replaces this with domain-specific classes, an evaluation set drawn from your data, a calibrated threshold tied to your cost of error, and monitoring for drift once it is live.
Custom classes, your accuracy target, your hardware constraints.