New data augmentation library released by Carvana winners

(Marco) #1

" … I am happy to announce that @AlBuslaev @creaf and I are open-sourcing a library for Image Augmentations. It is fast. It is diverse. Supports classification, segmentation, detection out of the box. Was used to win a number of #DeepLearning competitions."


do you plan to support training data (i.e. bounding box or object mask) augmentation as well?

(Marco) #3

It isn’t my library :slight_smile:
However it does support masked augmentations.

I am going to see if I can make a notebook and integrate it into fastai. It has a very nice pipelining mechanism.

(Charm) #4

It’s very cool! Look forward to supporting !

(Alexander Buslaev) #5

Hi, I’m the one of authors of this library, so feel free to ask questions if any.

Now we fully support classification and segmentation tasks and have basic (horizontal-vertical flip) support for object detection pipelines, but library is easily extensible and I also have plans to add more object detection augmentations also PRs are welcome.

As for fastai integration - we can help with this if needed.

(Cedric Chee) #6

Thank you for open sourcing this library.

The timing is just nice as I am currently learning and experimenting with lesson 14 Carvana image segmentation. I have been trying to push the results further. As I was looking for ways to do that, some ideas crossed my mind. One of them is do an assessment of the Carvana winners’ solutions to see what we can learn and apply to our own experiments. Then this came. I think it’s worth to try out the augmentation library.

Question: I have not dig into this library code. Does this library supports multi-threading and multi-processing for image resizing with OpenCV like what fastai library is doing behind the scene in the image transformation pipeline?