machine learning - What is switchable atrous convolution and why we use it?

Question

I was trying to understand a research paper called "DetectoRS: Detecting Objects with Recursive Feature Pyramid and Switchable Atrous Convolution". What I didn't understand is how the "Switchable Atrous Convolution" works and why we have used it? I know what Atrous/ Dilated convolution is, but what is Switchable here? How it is determined? I had spent days trying to grasp the concept before asking this question here.

Here are the links that I have collected and read from (might help you as well):

Official research paper at arxiv

Medium blog to get high-level overview

Python Implementation of Switchable Atrous Convolution (official GitHub repo)

I really value your time.

Thank you.

question from:https://stackoverflow.com/questions/65833638/what-is-switchable-atrous-convolution-and-why-we-use-it

Answer 1

I am answering my own question in the hope that it will be helpful for other people. SAC works like a soft switch, more like a mixing coefficient, which tells what information to take from both the atrous convolution (having different atrous rate) and mix them up. As "S" is dependent on 1x1 convolution has a trainable parameter, this helps the network to learn the optimal mixing coefficient. This is how our algorithm is looking twice at the image with the different receptive fields (different atrous rate) to capture important semantic level information which is important for object detection and semantic/ instance segmentation.

These images helped me a lot to unfold the information.