Week#3, Lab Image_segmentation_Unet_v2

For the record, I just want to say that the sentence “each pixel in every mask has been assigned a single integer probability that it belongs to a certain class” is proof that we are actually living in the future.

1 Like

Yes, the task of labelling data to create the training sets for Image Segmentation must be pretty costly and time-consuming. And of course you need a lot of labelled data to train a model like this.

I’d actually phrase that sentence a little differently: the label value is an integer that gives the “class” of the pixel that has the highest probability of being the correct identification of the object type of the object that the pixel belongs to. A probability value is not an integer, of course.

Definitely a better description… I almost pointed out the integer problem - but I had yet to confirm and thought: “well maybe they are converting to integer?”

My main point was just that when you step back and think about having the computational power (and RAM) to label every individual pixel of an image… it’s truly kind of staggering (at least to me).

Right! It’s amazing what the algorithm can do, but my point is there is a lot of human work that needs to happen in order to provide the training data that allows the algorithm to learn what it needs in order to perform that magic. There are some companies that are providing tools and services to do this kind of data labelling, but I have not really looked into the details there. My guess is that “There be dragons!” :scream_cat:

Ah - I see what you mean. Labeling this data does indeed sound like a nightmare. That wasn’t discussed much in the course content… I wonder if there are algorithms that could at least speed that process by Attempting to label things and having a human say “no”, “no”, etc. Still kind of a nightmare… but maybe less so.

1 Like