Custom dataset about image-segmentation-keras

[EvilPerfectionist]

@Zumbalamambo Hello, please look at the code in LoadBatches.py at line 46-49(

image-segmentation-keras/LoadBatches.py

Line 46 in 15a9b66

img = img[:, : , 0]

). "img = img[:, : , 0]" -> The blue channel is the first channel, so '0' in the equation is the blue channel. And in the following lines, you can see that the pixel value in blue channel represents the class labels.

[EvilPerfectionist]

@Zumbalamambo According to your first question, there are many methods to make annotation images. First, some large datasets provide the corresponding annotation images for raw image, such as PASCAL VOC 2012. If you cannot find the suitable dataset for your project, you can make your dataset by yourself, using either photoshop or some opensource software. You can search them on the internet.

[Zumbalamambo]

@EvilPerfectionist

This is how the annotation image will be if i use photoshop. But the masks are looking different.Its of light color. How do we do that?

[lktinh2018]

@EvilPerfectionist

This is how the annotation image will be if i use photoshop. But the masks are looking different.Its of light color. How do we do that?

We need to convert image with class value from 0->nClasses, because in load image function we assign label from zero->nClasses.
I will try with my dataset and give you my result if it's completed.

Best regards,
Tinh Lam