Geekly Articles each Day
📜 ⬆️ ⬇️

We apply the Deep Watershed Transform in the Kaggle Data Science Bowl 2018 competition.

We apply the Deep Watershed Transform in the Kaggle Data Science Bowl 2018 competition.


We present you the translation of the article by reference and the original dokerized code . This solution allows you to get about the top 100 on a private leaderboard at the second stage of the competition among the total number of participants in the area of ​​several thousand, using only one model on one fold without ensembles and without additional post-processing. Given the instability of the target metric in the competition, I believe that adding a few chips described below in principle can also greatly improve this result if you want to use this solution for your tasks.



solution pipeline description


Tldr


( Translator’s note - some terms are left as they are, that is, I’m not at all sure that there are adequate analogues in Russian. If you know of those, write in a comment - we’ll make some changes).


Each year, Caggle launches the Data Science Bowl competition. Last year it was pretty cool:



Unfortunately, when last year the competition started, I was not yet ready to take part in it. This year, after Google bought Kaggl, I began to notice some of the first "alarm bells" (a couple of "notes in the margins" - here and there ). Briefly - before the machine learning competition seemed to me mutually beneficial for both the community and the organizers of the competition, but now I am seeing some strange trend for the worse - it feels like the competitions turn into data marking exercises and / or awards become unattractive regarding the amount of effort that must be applied to participate normally (sort out / get to the top or give prizes / pump over).


Why I didn’t like the organization of this competition:



At first, when I opened the data, I generally did not want to participate, because their volume in megabytes did not inspire confidence at all. But then I examined them more closely and realized that the task here was in instance segmentation , which was a novelty for me. The task itself - instance segmentation is very interesting, despite the small size of the dataset. It is expected that you will not only create an exact binary mask of cells, as well as divide adherent cells (sorry, maybe there are nuclei, not cells, but judging by the markup, the organizers themselves are not sure about this either). On the other hand, the size of the dataset and the quality of the markup seemed a bit inadequate, especially considering the fact that the organizers of the competition, among other things, reported that some companies have similar datasets with terabytes of data .



Basic computer vision tasks. Here in the list, in theory, there should also be a classification of objects (the classic task is to find cats and dogs in the photo)


In this post, I will explain my approach to solving this problem. I will also share my pipeline inspired article Deep Watershed Transform for Instance Segmentation and tell you about other hikes and solutions, and also share my opinion on how such competitions should ideally be organized.


EDA or why ML is not magic


The training dataset contained about 600 images and the validation dataset - 65. The deferred test dataset from the second stage contained ~ 3000 images.
The images from the first stage had different resolutions - which in itself was a challenge - how would you build a universal pipeline for all of them?


 256x256 358 256x320 112 520x696 96 360x360 91 512x640 21 1024x1024 16 260x347 9 512x680 8 603x1272 6 524x348 4 519x253 4 520x348 4 519x162 2 519x161 2 1040x1388 1 390x239 1

There were about three clusters among the training data that are easy to find with K-means:



This was the main reason why converting RGB images to black and white helped on a public leaderboard.



Black images



Different variations of cores in shape, color, size


Visual viewing of a test dataset with three thousand pictures showed that 50 +% of these pictures have nothing to do with the training dataset, which caused a lot of controversy and resentment from the community. So you participate in the competition for "thank you", spend time, optimize the model, and bang and get 3000 pictures that do not look like training data? It is clear that the goals may be different (including preventing manual marking on between the stages of the competition) - but this can be done much less clumsy.


Some noteworthy files from the test dataset:



I would suggest that the little thing in the background is the core



Honestly no idea what's



Looks like muscle. Once again, are these white pieces cores or what?



Night sky ... Is it a core or just noise?


Deep watershed transform


If you do not know what it is, then go here . Intuitively, the watershed method is fairly simple - it turns your image into a negative "mountain landscape" (height = intensity of pixels / masks) and fills the pools of the markers of your choice with water until the pools connect. You can find a bunch of tutorials with OpenCV or skimage , but all of them usually ask questions like this:



Deep Watershed Transform (DWT) will help us solve some of these problems.



The main motivation of the original work



The idea is for CNN to take into account two things - single vectors indicating the boundaries and energy level (height of the mountains)


In practice, if you simply apply a WT (Watershed Transform), you will most likely get a very parcel segmentation. The intuition behind DWT is such - we must teach CNN to find the "mountain landscape" for us.


The authors of the original article used 2 separate CNG VGG-type to obtain:



In practice, you can use one network or just train several smaller end-to-end networks . In my case, I played with networks that gave out:



Then you need a little magic to combine it all and voila, you have the "energy". I didn’t experiment a lot with architecture, but Dmitro (the author of the solution above) then told me that he didn’t get better results when adding the second CNN.


For me, the best post processing ( energy_baseline function by reference) was the following algorithm:




One of the best examples - the net was able to clearly separate the stuck cores



The gradients learned are not suitable for use as a watershed.


Sometimes a direct search for nucleus centers also worked, but overall, this did not help improve speed.


Other approaches


Personally, I would divide the possible approaches of this task into 4 categories:



For me, in favor of choosing DWT + UNet, this is a solution, where you don’t have to bother too much, because it is simple (you can simply feed layers of energy as additional channels for the mask) and then easily shift the work to other tasks. I also really like the recurrent extension UNet, but I didn’t have enough time to try it.


In the case of recurrent UNet, there are three new components in essence compared to regular UNet:



It all seems stunning at first, but I will definitely try it in the future. Although, this method combines two memory-hungry architectures — the RNN and the encoder-decoder network — which may not be practical in actual use outside of small datasets and permissions.



ConvLSTM Layer Description



Recurrent Unet Architecture


My pipeline


You can find the details here , but my approach is as follows:




Whole pipeline


If you want to greatly improve the result of this pipeline, then you need to replace the VGG-16 encoder with Resnet152 - according to the participants of the competition, this encoder behaved much more stable on delayed validation. Also replacing softmax with sigmoid as the last activation function gives less blurred borders.


And now about how, in theory, such competitions should be organized


In short, SpaceNet from this point of view was almost perfect if you discount the annoying moments of the TopCoder platform:



Thanks


As always, many thanks to Dmytro for fruitful conversations and tips!


References:



')

Source: https://habr.com/ru/post/354040/

More articles:


All Articles