[GroupSampler] How can GroupSampler bring any benefit?

[songyuc]

Hi, guys,
I am learning MMDet today, and I found it in a Chinese tutorial,

Specifically, it says, "由于 Dataset 中输出的图片大小不一样，为了尽可能减少后续组成 batch 时 pad 的像素个数，MMDetection 引入了分组采样器 GroupSampler 和 DistributedGroupSampler，相当于在 RandomSampler 基础上额外新增了根据图片宽高比进行 group 功能";
English version: Since the output images in the Dataset are of different sizes, in order to minimize the number of pixels in the pad for subsequent batch composition, MMDetection introduces GroupSampler and DistributedGroupSampler, which are equivalent to adding an additional group function based on the aspect ratio of the images on top of RandomSampler.

However, as far as we know, most current detetors use a fixed size for input image. As a result, it seems the point of "minimize the number of pixels" does not make sense. Unless there is any other detectors that use dynamic sizes, so, is there a detector use different input sizes within different batches?

Your answer and guide will be appreciated!