Hello !

I have a general question on U-Net Lab.
How the mask dataset was produced out of the image dataset and was it necessary ?
Also I did not understand which role did it play in all the story ?


The masks are the “labels” on the data, right? You need those in order to train the model. I don’t specifically know how they were created, but my guess is that someone had to manually go through and do the work of labelling the regions. Sounds like a lot of work.

1 Like

Hello @Paul !

I mean there are two folders with images: data/CameraMask and data/CameraRGB.
The one is in color, and the other is really dark like watching through the IR filter.
I was wondering why do we need it and where was it used during the training?

You’re just not rendering the mask images correctly. They show you how to do it in the notebook and they show you what they look like. The problem is that they only have a single color value for each pixel to save space. If you clone the pixel values for the other two “colors”, then they render correctly.

As to how they are used, it is all laid out in the notebook. They are the “ground truth” values corresponding to each input image. The goal of the model is to output such masks and then we train the model to try to produce outputs that are as similar as possible to the “ground truth” mask images.

so basically what is called segmentation in the beginning of the lab when showing image and corresponding mask is a red channel information of the masked image.