Decided to start working in a more difficult competition for image recognition

Deep Learning
1

Here is the competition: HuBMAP + HPA - Hacking the Human Body | Kaggle

Now this competition is quite difficult. I decided to start small and try to predict the organ where the image is from.

I created this easy dls

dls = DataBlock(
  blocks=(ImageBlock(cls=PILImageBW), CategoryBlock),  
  getters=[ColReader('path_id'),   # image input
           ColReader('organ')],
  splitter = RandomSplitter(valid_pct=0.2, seed=42),
  item_tfms = Resize((640,480), method='squish'),
  batch_tfms= aug_transforms(size=(288, 224), min_scale=0.75)
).dataloaders(df)

Where I added to the df the path where the image is store so that I can use the getter attribute to look for the image and the label.

Then I build a learner.

learn = vision_learner(dls, resnet50, metrics=error_rate)

image
image536×736 59.6 KB

The result was quite good. However, I’m stuck. I really don’t understand how to build the actual label we need to predict. I think that maybe this is an image segmentation problem because you get a JSON file with a lot of different coordinates.

1 Like
2

You don’t mention the competition’s discussion site: HuBMAP discussion
That seems the best place to ask competition specific questions.

1 Like
3

I did not look at all details (I would have to join the competition), but yes, this is a segmentation problem. The main difficulties are the variety of image sources, sizes, organs.

1 Like