Surprisingly, the accuracy of the CNN I trained only on a half of available images does not distinguish significantly from the accuracy of the CNN I trained on all data (with respect to standard deviations). In other words, the CNN I (performing alone) obtains the same performance as the baseline model by Jia and Cristianini [11] with the training set of only 225, 000 images, which is almost 20 times smaller than the training set used by Jia and Cristianini