What are the advantages of the super-resolution module (ARRD module in paper)?

[anonymoussss]

In my opinion, using only Down-sampling Degradation Transformations allowed the detection network (both encoder and decoder) to learn the ' intrinsic equivariant representation against various resolutions and degradation'，isn't it?

What are the advantages of the ARRD module over the trick 'multi-scale training' commonly used in detection training? It seems to just add a supervision.

Is there a mathematical explanation?

[cuiziteng]

Many thanks for the valuable question, we use ARRD to constrain that Manifold with Scale invariance, If there is no ARRD, it is a simple data argumentation, and there is no ability to scale invariance.

For more math explanation about the scale invariance (etc.), you could refer to AET and this paper, hope this could solve your question.