How to divide the dataset?

[NicholasIrving]

How to divide the training, validating and testing set?
I read the paper Learning Rotation-Invariant Convolutional Neural Networks for Object Detection in VHR Optical Remote Sensing mages , it said that training set contains 120 images, validating set contains 120 images and testing set contains 410 images. But I found that if I divide 0-120 as training set, 121-240 as validating set and 241-650 as testing set, the number of classes are not balance.
In training set:

In validating set:

In testing set: