hi , i’m trying to do a segmentation on satellite images but the details are not promising . is there anyway to train a model that can detect smaller details . right now i’m using deeplabV3 with resnet50 as backbone
You haven’t said much about your use case. In particular, are you
performing semantic segmentation (where you wish to label each
pixel with the class to which it belongs)? Or are you performing
instance segmentation (where you label each pixel with the specific
instance* of one or more classes of objects to which it belongs)?
Also, what is the structure of your images? Do, for example, the pixels
in the lower left of your image provide any significant information about
to which classes the pixels in the upper right of your image belong?
Without knowing anything more, I would suggest that you try U-Net, but
how deep you want your U-Net to be (which is to say how large a field
of view you want it to have) will depend on your specific use case.
*) An example of instance segmentation: Let’s say you have images
that contain bicycles and cars (and people and trees and fire hydrants,
but you don’t care about those). Any pixel that is not part of a bicycle
or car is a background pixel. Let’s say that some specific image contains
two bicycles and three cars. Instance segmentation is used to label each
pixel as being “background,” “bicycle-1,” bicycle-2," “car-1,” “car-2,” or
“car-3.” (By way of contrast, semantic segmentation would label each
pixel as being “background,” “bicycle,” or “car,” but wouldn’t tell you
which specific bicycle or car the pixel is in.)