Apply to RGB pictures

[giovanni-turra]

I would apply U-Net to RGB pictures (input). Could I understand if is it possible with a small dataset? Is it possible with U-Net and this implementation?

[jocicmarko]

Yes, it is possible, just make sure you specify that input has 3 channels:

inputs = Input((img_rows, img_cols, 3))

[giovanni-turra]

Thanks Marko.

But, from a structure point of view, U-Net apply an averange across color axis after input step?

Or is it able to combine information from single channel to increase segmentation phase?

I ask you this because from U-Net paper I think this part it is not covered.

[titusaj]

If changing over to color image segmentation I took the approach where the train_mask images are also converted to three channel images for ease of input into the model.

However I get reshape/incompatible shape error when run.

The model runs fine at the input level (ie. accepts images into model).

How would the model itself have to change to allow for three channel input/convolution.

[namsu-kim-unsw]

I tried making mine RGB and ran it but it outputs a completely black picture

Edit: I think the reason for this was because of the number of epochs - it was too low. With a higher number of epochs, it did generate an image but the dice is super low :( around 5%

[matsujju]

@jocicmarko Where to input images and mask in the data.py file if I have image data outside kaggle data?