Assymetric Kernel Size

While the symmetric f x f kernel seems obvious, how about an asymmetric kernel of size f1 x f2?

Do you believe this would provide some benefits?

Not able to get the intuition right as to why an symmetric filter is preferred … hence trying to reach out to folks if anyone has some convincing inputs.

Else I am attempting to implement an asymmetric filter and uncover why it isn’t so popular.

A brief internet search turned up this:
https://www.researchgate.net/post/Why-are-the-kernels-used-in-CNNs-square-matrix

I think Hassan’s answer in that thread is the most useful.

Thank you for sharing the thread. I did go through the trail however am still looking for some convincing factual experimental results.

Nevertheless, I am working on the same to convince myself of all the pros mentioned in the thread.

Thanks again.