Segmentation algorithm not giving correct output

Hi ,
I am evaluating segmentation models with my own data. Only FCN produces the right segmented output for test images, but other segmentation models (such as U-Net and SegNet) produced really odd results. The segmented pixels are dispersed throughout the whole image instead of the targeted region. Even if I evaluate my trained model using the training data.
What may be the reason behind that?
This is segmented output by SegNet.

This is segmented output by FCN.

[sudobash]

Your images are not visible.

Answer 1

[Birju Patel]

Generally, FCN, U-Net, and SegNet are different architectures that require their own set of training options to produce optimal results. You cannot assume they will all converge to the same results.

For instance, U-Net does not use a pre-trained backbone so it can take longer to train compared to FCN, which uses a VGG-16 image net pretrained backbone.

[Shoaib Ali]

Thank you for your answer.

You mean I have to train each network with diffferent hyperparameters to figure out the best values of parameters for each model?