why fast rcnn's input is different from the training prototxt

[milk-bottle-liyu]

I am a green hand and sorry to bother you about this closed project, but I'm quite confused about the test.prototxt of the VGG16 model, in which the input is
input: "data"
input_shape {
dim: 1
dim: 3
dim: 224
dim: 224
}

input: "rois"
input_shape {
dim: 1 # to be changed on-the-fly to num ROIs
dim: 5 # [batch ind, x1, y1, x2, y2] zero-based indexing
}

is this natively supported by caffe or Mr. RGB change the source code?