Semantic segmentation with overlapping patches

I am using Deeplabv3+ to perform semantic segmentation in an image but the result shows strong discontinuities due to patch-wise predictions. Is there a way of using overlapping patches that are aggregated (or averaged) to produce the output? I tried to use a code showed in this example. However the output has still the patchy pattern. Any sollutions?

[Mahesh Taparia]

Hi

Can you specify the input which you are passing? I think the network give the pixel wise output, how it became discontinuous?

[Matheus Ferreira]

Because I am working with remote sensing images I should pass image tiles that are merged together after prediction. I could solve the problem using 'blockproc' and defining the 'BorderSize' parameter according to the amount of overlap desired.

Answer 1

[Mahesh Taparia]

Hi

In this case, you can use 'blockproc' function. Create the image block and pass it to network. For more information, you can refer this documentation.