gcForest for images

[dangmanhtruong1995]

Hi.
In the gcForest paper the author states that three sliding windows are used: d/16, d/8, d/4 are used, where d is the number of features. However I don't understand about the case for images.
For example, with an image of size 472 * 677, for the case of d/16 would it be of size: (472 / 16) * (677 / 16), that would be a window size of (30, 43). The number of instances for one image would be: (472 - 30 + 1) * (677 - 43 + 1) = 281305. So the number of pixels would be: 281305 * 30 * 43 = 362883450. Compared to original: 472 * 677 = 319544 pixels. So the dataset would be bigger by 362883450 / 319544 = 1135.63 times.
It seems to me that this would make the new dataset too big.
I wonder if this is correct and what the author does to mitigate this?
Thank you very much.
Truong.