thx for the video. Some of the masks in the label folder consist only of background, there is no class to segment. Because when we patchify the big image at the beginning, there is nothing to label in some of the patches. Don't we need to extract these images before training them?
Hey, good question! Let's discuss. Your vision task/type of classification would determine whether you (1) should use sigmoid or softmax activation function after your final convolutional layer, (2) Binary Cross Entropy or Categorical Cross Entropy for your loss function, (3) Do you put 1 or 2, or other number in n_class? Etc. I presume your vision task is binary classification as well? If I am not mistaken, my model still perform well despite being trained with some of patches with no class. Do you have any resource to back your theory whether including those background-only patches in the training would introduce more harm than good e.g. noise thing? Perhaps if you dont want them to be on your patch-folder/model development, you need to subset your images-to-be-patchified first, so you would have "cleaner" images. As of automatic removal of background-only patches in the folder, I do think there is a way to do that. I dont know yet how. Again, good question! Stay curious!
@@geografif yes exactly , i have a binary class segmentation task. actually i don't have any theorem whether including background only patches affect the result seriously. I hope to get an idea by trying both ways and comparing their accuracy. While doing this, I plan to select and eliminate background-only images by checking them one by one. Obviously automating this operation isn't a simple if-else condition. Structures where label is uniform need to be parsed.
@@geografif By the way, when I removed the background only patches (which decreased from 570 images to 330), my accuracy decreased from 82 to 69. This means that it accepts background only patches as 100% correct.
Thanks for the update. Anyway, have you tried to apply your model to the unseen dataset where you dont have background-only chips? That's a way to measure accuracy and see how your model perform outside the 'seen' datasets used in the training phase.
thx for the video. Some of the masks in the label folder consist only of background, there is no class to segment. Because when we patchify the big image at the beginning, there is nothing to label in some of the patches. Don't we need to extract these images before training them?
Hey, good question! Let's discuss.
Your vision task/type of classification would determine whether you (1) should use sigmoid or softmax activation function after your final convolutional layer, (2) Binary Cross Entropy or Categorical Cross Entropy for your loss function, (3) Do you put 1 or 2, or other number in n_class? Etc.
I presume your vision task is binary classification as well?
If I am not mistaken, my model still perform well despite being trained with some of patches with no class.
Do you have any resource to back your theory whether including those background-only patches in the training would introduce more harm than good e.g. noise thing?
Perhaps if you dont want them to be on your patch-folder/model development, you need to subset your images-to-be-patchified first, so you would have "cleaner" images.
As of automatic removal of background-only patches in the folder, I do think there is a way to do that. I dont know yet how.
Again, good question! Stay curious!
@@geografif yes exactly , i have a binary class segmentation task. actually i don't have any theorem whether including background only patches affect the result seriously. I hope to get an idea by trying both ways and comparing their accuracy. While doing this, I plan to select and eliminate background-only images by checking them one by one. Obviously automating this operation isn't a simple if-else condition. Structures where label is uniform need to be parsed.
@@geografif By the way, when I removed the background only patches (which decreased from 570 images to 330), my accuracy decreased from 82 to 69. This means that it accepts background only patches as 100% correct.
Thanks for the update.
Anyway, have you tried to apply your model to the unseen dataset where you dont have background-only chips?
That's a way to measure accuracy and see how your model perform outside the 'seen' datasets used in the training phase.
@@geografif I'm dealing with exactly this right now, I'll share the results.