2 posts • Page 1 of 1
I don't understand using the train_32x32.mat database for the CNN course without preprocessing to remove left and right edges. A lot of the 32x32 pictures contain more than one number, but the designated label is the digit which has been placed in the center of the image even if the image contains other perfectly legible digits. Since CNNs are supposed to be insensitive to translations, a picture which contains, for example "318" is labelled as "1" because the "1" is in the center but both "3" and "8" are perfectly reasonable labels (see class 1 image 3). There are also a fair number of images with the central digit less legible than the digit beside it and in that case the "best-looking" digit is not the class label (see class 7 image 3).
- Screenshot at 2019-05-29 16-49-06.png (195.79 KiB) Viewed 318 times