Skip to content

Questions about the size of image (SwinUNet) #16

@waner1998

Description

@waner1998

Dear Author,

Hello! After reading several papers, I finally came across your article. I would be extremely grateful if you could help me with a few questions.

In your paper, you mentioned resizing the input images to 352×352 and also reported results using SwinUNet. However, I found it difficult to reproduce good results (e.g., on CVC-300, the Dice score is only around 0.44) based on the official SwinUNet code. I'm wondering if this issue might be due to the pretrained weights—since the available SwinUNet pretrained models are only for 224×224 or 384×384 input sizes, and cannot be directly loaded for a 352×352 model. Could you kindly share how you addressed this issue?

Additionally, I have a minor question: you mentioned using multi-scale data augmentation with scales {0.75, 1.0, 1.25}. How did you handle the impact of this augmentation on SwinUNet? Since the input image size would vary due to this operation, it could cause errors if the input size no longer matches the model's initialized input dimensions.

Thank you very much in advance!

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions