I am hosting a Kaggle competition - Thanks to fast.ai!

(Stephan Rasp) #1

Hello to everyone in the fast.ai community,

I would like to share a Kaggle competition that I am co-hosting. The competition is called “Understanding Clouds from Satellite Images” and has just launched with a total of $10,000 prize money. It is an image segmentation challenge but with a twist.

The task of the competition is to detect four different cloud patterns (Sugar, Flower, Fish and Gravel) from satellite images. (If you want to find out why we care about this sort of thing, check out the arXiv paper or the blog post below.) The first difference to typical segmentation tasks like Carvana is that the classes are subjective. Further, they do not have clear boundaries. The way the ground truth labels were created is through a crowd-sourcing activity (try it out here) where around 70 scientists sat down for a day and labeled 30,000 images. Since there is no definite truth, the challenge in this competition is to build a model that agrees with the human average.

The reason I am sharing this here is that, for me personally, this Kaggle competition is the latest step on a machine learning journey that started with fast.ai about two years ago. I was a PhD student in meteorology at the time, a little fed up with my research, and curious to learn more about this “machine learning” everyone was talking about. So after doing Andrew Ng’s original ML course on Coursera, I discovered the first version of fast.ai. I started it and immediately became hooked. I don’t think I ever learned as much in such a short amount of time as I did going through the fast.ai lectures.

In the two years since, I’ve worked on various projects that try to improve weather and climate predictions using ML, this cloud classification project among them. I love the work that I do and I would like to take this opportunity to thank @rachel and @jeremy for making this possible.

If you want to find out more about the competition and data, check out the blog I wrote and the arXiv paper.

I hope that many of you will participate in the Kaggle competition. In fact, for the arXiv paper I used the fastai library to create the segmentation model and it was pretty easy, so there are no excuses :wink:

22 Likes

#2

Looks interesting! Thank you for hosting and sharing.

0 Likes

(Biju) #3

Great effort.All the very best.Keep informing.

0 Likes

(Ilia) #4

Really interesting competition! I am going to participate. Never worked on segmentation tasks, and this one seems to be accessible even for people with not so much experience in the field.

0 Likes

(Cesar Calderon Muro) #5

I was trying to work on this competition but I can’t get the datablock api working to load the data. Can somebody share if they have been able to load the data with fastai’s datablock API?

Thanks!

0 Likes

(RobG) #6

In this dataset, masks for classes may overlap. By default fastai wants to see one class per pixel, in one channel. This is likely causing your probem. You’ll need to alter fastai to handle one channel per class. By kaggle rules, any such code discussion and examples should be discussed on kaggle forums. But if you want to talk in generic terms without reference to a challenge, this thread is tackling the same problem :slight_smile: Create databunch with multiple segmentation mask as label

3 Likes

(Sergio) #7

I created a follow-up notebook for this competition using fast.ai. It relies on the previous implementation but also adding some post-processing to achieve the best accuracy by optimizing threshold selection. I hope you find it useful :smiley:

https://www.kaggle.com/sramirez/cloud-segmentation-with-fast-ai-post-processing?scriptVersionId=23962088

1 Like