Tools for image data augmentation

[catar...@gmail.com]

Hi everyone,

I am looking for open source tools for image data augmentation (rotate images, change contrast,...).

I came across the python library albumentations (https://albumentations.ai/) and I was wondering if anyone in the group has used it before and recommends this or any other open source tools for this job?

Thanks in advance! 

Catarina

[Zhongliang Zhou]

Hi Catarina,

For me, I used torchvision's augmentation library: https://pytorch.org/vision/main/transforms.html. It works well if you want to incorporate the data augmentation process with some deep learning models that are written in pytorch.

Best,

Zhongliang

[Joar Nykvist]

Hi Catarina!

Sorry I did not see this earlier. I have used an algorithm called RandAugment which is implemented in pytorch: https://pytorch.org/vision/main/generated/torchvision.transforms.RandAugment.html.  

Paper: https://arxiv.org/abs/1909.13719

Brief explanation video https://www.youtube.com/watch?v=Zzt9i3gDueE. 

All the best,

Joar Nykvist

[Catarina Silva]

Thanks for your replies Zhongliang and Joar!

I ended up using the albumentations library and it works quite well in our pipeline. It has a lot of supporting documentation and a lot of functions for image transformations (along with transformations of the actual annotations - bounding boxes, segmentation, keypoints...).

Xabier Lekunberri also pointed this tutorial out from OpenCV, which he used to change contrast and brightness, and might be useful to other people in the group: 

https://docs.opencv.org/3.4/d3/dc1/tutorial_basic_linear_transform.html

Cheers,

Catarina