Live detection

What is actually running

• Weights are a YOLO26s ONNX export at imgsz=960, opset=12, trained at 1280 on the gun dataset.
• The export is end-to-end: NMS is inside the graph, so the model emits a fixed [1, 300, 6] tensor and the page only has to threshold it.
• Frames are letterboxed to 960×960 with grey padding, matching the training-time preprocessing — aspect ratio is preserved, so boxes are not skewed.
• One class only. Index 0 is gun; there is no COCO label table.
• The 38 MB model is cached by the browser after the first load, so a revisit starts immediately.
• Export composites the frame and the overlay onto a second canvas and records it with MediaRecorder — the boxes are burned in, and nothing leaves the browser. It asks for MP4 first and falls back to WebM, and the file is named for whichever container the recorder actually produced. Older Safari has no MediaRecorder at all, so the button is disabled there.
• Picked video files are capped at 30 seconds. The camera feed has no such limit — it is a continuous stream with no duration to check.