about multi-scale traing/testing

[foralliance]

@YuwenXiong
@oh233

For example, 5 scales: [A, B, C, D, E, F]

For multi-scale training,
in fact was only trained for 1 time, but each scale is random? In other words, multi-scale training does not mean that every scale is trained once?
that's my understanding, i'm not sure it is right.

For multi-scale testing,
According to the paper Deep residual learning for image recognition and Object detection networks on convolutional feature maps, multi-scale testing refers to: each test, select two adjacent scales for feature extraction, and feature fusion, and then test.
my question is：
is carried out 5 times (near the two scales each test (such as: AB, BC, CD, DE, EF)), or just 1 times (randomly selected one set of neighboring two scales)?