Comment by theamk

Sure there is training - most few practical algorithms have dozens of tunable parameters - bucket size, thresholds, camera settings, image normalization settings and so on. It may not be 175 billion weights, but this still needs plenty of training data.

I've participated in hobby robot competition in the past, which required simple-sounding vision part: find a bright orange object on a green grass in bright sunlight, and very roughly estimate distance. We had to get 200+ training images and manually label each of them to get any sort of decent performance.