What is the expected output of the submission?

In the suggested baseline, there is a 3-class classification model for each pixel. I would like to ask to clarify the expected output of the submission.

Do we have to develop a Semantic Segmentation model which predicts 3-class on each pixel in the image?

or

Do we have to develop a Instance Segmentation model which seperately segment each cell with binary label classes with unknown number of instances in the input?

Thank you!

This is an instance segmentation task. The expected output of your model is the same as the groud-truth label.

I am sorry but things are a bit confusing for me still, so I will try and break down my understanding of the task for myself and others maybe :

The labels provided in the training set are grayscale images with values in the range 0-2 (0=background, 1=cell body, 2=cell boundary) The model provided in the baseline solution is one for Semantic Segmentation which does instance segmentation using morphopgy processing on the predictions. The output from the predict.py file in baseline- is a mask with grayscale values between 0 and "N" (0 = background, 1=cell #1, 2= cell#2, ... N = cell#N)

So in my opinion the OP was correct, but I am confused when you say "the expected output of your model is same as ground truth label" coz the provided predict.py doesn't do that, rather it outputs labels like I described above.

Apologies for raising this question again but it seemed quite valid to me coz if we are gonna predict same thing as the provided GT labels that would mean accurately detecting the cell boundary class as well! and since they account for so little pixels in the image it will need to be tackled in it's own way but if we are just gonna do instance segmentation then the cell boundary labels don't seem of much importance to me for that task as evident from the output of your predict.py

Hi ,

Thanks for raising the questions.

Conclusion first: This is an instance segmentation task and you need to give each cell an individual label.

You mentioned that

The labels provided in the training set are grayscale images with values in the range 0-2 (0=background, 1=cell body, 2=cell boundary)

This is wrong. Could you please dobule-check the ground truth? I'm sure they are instance cell labels (1,2,3,...) rather than cell-body/boundary.

In the baseline, we formulated the instance segmentation task as a three-class segmentation task. Then, we labeled each individual mask to obtain the instance mask, which is motivated by this work https://www.nature.com/articles/s41592-018-0261-2

However, it's not the unique solution for cell segmentation. We provided multiple papers in github, which can obtain better performance. https://github.com/JunMa11/NeurIPS-CellSeg#limitations-and-potential-improvements

Please feel free to raise any questions:)

Massive thank you for clearing it up!!

And yes you are right! I was looking at the GT labels after the preprocessing that was done by the baseline segmentation model from the repo :') and now when I checked the "RAW" (or lets say unprocessed) GT labels they are as you described.

Maybe it would be nice for new comers to the competetion if this detail about your baseline approach was also in the repo, but otherwise I am thankful to you for clearing this up.

Cheers!

you are welcome:)