search

zhixuhao / unet

unet for image segmentation
MIT License
4.6k stars 2k forks source link

Different results on different dataset #160

Open sabbban opened 4 years ago

sabbban commented 4 years ago

Hi When I ran this network on the ISBI dataset (same images as in this repo) the results are as expected to be - a binary results. When I ran it on different dataset images of colon biopsy I'm getting kind of grey scale images.

From the ISBI dataset test and predict images: https://ibb.co/tJhRzst

From the colon biopsy dataset results: test and predict images: https://ibb.co/jJXZthY

Label data of the colon biopsy dataset: https://ibb.co/BwjxKLZ

Any intuition why on the ISBI I'm getting a binary classification and on the daataset that I have I'm getting kind of grey scale transformation of test images ? The network didn't changed.

Thanks in advance

aedirn commented 4 years ago

You can also get grayscale blurry results if you train this network an insufficient amount on the example data it comes with. Change steps_per_epoch=300 to steps_per_epoch=100, and you'll get hazy grayscale results.

The output of the model's prediction is a float64 in range [0,1], and I guess with your data (and undertrained on the repo's example data) it isn't fully "confident" in all its pixel classifications so gives them a value less than 1.0.