Train Segformer with custom dataset

[pnchuyen]

Hello, thank you for your implementation. Have you tried to train the segformer model with any dataset. How was it in compared to U-Net model? Thank you and look forward to your answer!

[ACSEkevin]

Hello, thank you for your implementation. Have you tried to train the segformer model with any dataset. How was it in compared to U-Net model? Thank you and look forward to your answer!

Hi, Thanks for the comment.

Actually models like U-Net is not suitable for comparing Segformer as they are designed for different purposes.
U-Net which has a direct modelling for both encoder and decoder, is a simple-structured model and might not be able to learn complicated representations.
Whereas, models like PSPNet, Deeplab and other Self-Attention-based segmentation models are more capable for capturing complex features, context etc. because of their huge capacity, and modelling strategies which give them priors or some constraints so that guide them to learn better representations.

Therefore, U-Net might be suitable to learn from Datasets like medical images e.g. abnomality segmentation, while Segformer can be applied to datasets that have latent comlicated relationships e.g. Cityscapes ADE, COCO seg.

Hope you find it helpful.

[pnchuyen]

Thank you for your response! It is really helpful for me. Did you try to train the segformer model with above dataset? How was it?. Keep up the good work!

[ACSEkevin]

Hi,

Sorry that I respond late I've been busy with my work.
Please refer to the original paper which provides all the results.
Hope you find it helpful :)