MetalDetector
GoogLeNet on iOS demo.
Note: this demo was implemented before TensorFlow, PyTorch or Core ML existed. It's preserved for historical purposes only.
This is a basic demo of GoogLeNet from Caffe Zoo working on iOS. It takes the input from the camera and prints the top prediction. The speed is 1 FPS on iPhone 6S, and 3 seconds per frame on iPhone 6.
Two known bugs:
-
You need to hold the phone so that the round button is on the right. Otherwise, the network will get a rotated image, and the classification will likely miss.
-
There's a rounding bug right now (likely, in Convolution layers). While the network gives answers which are within ~3% of Caffe output, it's slightly worse than a real GoogLeNet taken from the Caffe Zoo.
Also, please, be aware that ImageNet classes are weird, not so many real world things could be detected (but it knows about 300 breeds of dogs, whee!)
The code was generated with Goffe, Caffe-to-{Swift+Metal} generator.