Have you considered using CapsNets? about dexined HOT 6 OPEN

There are a LOT of things I don't know about, so no need to be sorry. :)

The hope is it can reduce the number of parameters and make training quicker.

That is the objective, but before I want to generalize DL model. Train just one time but evaluate in the whole edge detection dataset.

from dexined.

The capsule rabbit hole has led me to these papers, which may be of interest to you:

http://sciencewise.info/articles/1804.04241/

http://sciencewise.info/articles/1907.13327/

http://sciencewise.info/articles/1912.08367/

The capsule version is delayed for now. :)

I will check it thanks for sharing

from dexined.

See: https://arxiv.org/abs/1710.09829v2 and https://www.oreilly.com/content/introducing-capsule-networks/

They are supposed to provide better results with fewer training examples.

I'm going to translate your updated TF2 version to a CapsNet and see if it makes a difference.

Hey I did not know about this paper sorry :( I am going to check this weekend

from dexined.

There are a LOT of things I don't know about, so no need to be sorry. :)

The hope is it can reduce the number of parameters and make training quicker.

from dexined.

I should have an initial prototype of the capsule version ready Monday or Tuesday. I'll then train on the original images in both cpu-only mode and GPU in Colab, test against everything.

It can be updated as needed.

from dexined.

The capsule rabbit hole has led me to these papers, which may be of interest to you:

http://sciencewise.info/articles/1804.04241/

http://sciencewise.info/articles/1907.13327/

http://sciencewise.info/articles/1912.08367/

The capsule version is delayed for now. :)

from dexined.