Frame buffers beat photos

The first version of the Lorcana scanner had a shutter button. You lined up a card, tapped, waited, got a result. About three seconds a card, and your thumb hurt after a playset.

The second version never captures a photo at all.

Read the buffer, skip the capture

Recognition runs on whatever is already flowing through the capture pipeline:

func captureOutput(_ output: AVCaptureOutput,
                   didOutput sampleBuffer: CMSampleBuffer,
                   from connection: AVCaptureConnection) {
    guard let pixels = CMSampleBufferGetImageBuffer(sampleBuffer) else { return }
    recognizer.submit(pixels)   // no still capture, no pipeline flush
}

The whole speed win comes from deleting the capture step:

• A still capture reconfigures the camera, which costs a pipeline flush and a couple of hundred milliseconds you can feel.
• The buffer is already there at sixty frames a second.
• You were throwing it away to take a worse version of the same image.

Any single frame is a bad frame

Motion blur, glare off the sleeve, a thumb over the set symbol. The fix is to never trust one frame:

Debounce on agreement, not on time. Three frames that say the same thing beat one frame that says it confidently.

Which you should have been doing anyway, shutter or not.