Random scaling in image augmentation

[vincentfung13]

Hi all,

I am new to mxnet (to deep learning in general), and in my project I am in a situation where I need to do some scaling on the images for augmentation.

I was using the provided ImageAugmenter class for this - setting the min_random_scale and max_random_scale parameters when initialising the ImageRecordIter. However, after some digging into the source code I found what the augmenter actually does is to scale the image then crop it to fit the predefined data_shape. What I want is to randomly scale the image while retaining its content, which means an image is first scaled, then resized again to the predefined data_shape.

I managed to change the original ImageAugmenter class so it now does what I wanted, but I am not sure if it is the best place to put the changes in (potential merge conflicts if something new is added to the augmenter in the future). I can think of implementing my own augmenter but it looks like I will need to implement a new ImageRecordIter and a ImageIOParser as well, I am not sure if the effort is worth it if all I want is just to add a new augmentation approach.

Can someone point me to a right direction? Any help is much appreciated. :)

Cheers,
Vincent

[tqchen]

there is a recent refactor that enables you to plugin your own augmenter by registering it.

see #1956

So now you can build your augmenter class in a new file, register it as aug_default.cc and switch to your augmenter by setting aug_seq in ImageRecordIter

[vincentfung13]

Thank you for making the refactor, this looks much more feasible now. :)

[shipengai]

Hi, how do you make your own augmenter.Can you share it with me ?

[TccccD]

Hi, @vincentfung13 @shipeng-uestc @tqchen
This seems a bit confusing, is there a simple example of using aug_seq? I want to use like"random crop - > resize", this resize method stretches the image to a square.

[thomelane]

Hi @TccccD,

I've just replied to a similar post here which hopefully clarifies things a little more.

Aside from implementing your own augmenter and using in aug_seq, you should take a look at Gluon's Transforms. You can Compose augmentations in the order you need.