The point is that you have two different distributions of images: 900k from the internet and 100k from your car’s front facing camera. It would be a mistake to use only one distribution for any of the three datasets (training, dev and test), right? So your solution would mean that the dev and test sets only contain images from your car’s camera and none of the internet images. That can’t possibly be the best answer. Which of the remaining possibilities sound like they would satisfy the requirement?
1 Like