How to process the training data?

[buffazombie]

Thank you for the great work! I’d like to ask about the preprocessing steps applied to the input LDR images during training. Specifically, what kind of processing (e.g., normalization) was done before feeding the images into the model?

[cecret3350]

Thank you for your recognition of our work.
The over-exposure image is normalized to [-1, 1] to meet the requirement of VAE. We normalize the under-exposure image to [0, 1], mask with a pseudo occlusion mask and then extract the structure and color information.
We also apply widely-used augmentation techniques (random crop, resize, rotate, flip) during the training process.

[buffazombie]

Thank you for your patient answer, it inspired me a lot!

[buffazombie]

非常感谢大佬的回答！大佬我还有一个小问题，因为我是想用RAFT做一下对齐来看一下效果的。所以，在图像处理之后输入进mef函数的img1和img2原本位深是多少的呀，另外在输入进去是不是也做了归一化到01之间？ 原始邮件 发件人：Zixuan Chen ***@***.***> 发件时间：2025年6月22日 16:47 收件人：OpenImagingLab/UltraFusion ***@***.***> 抄送：buffazombie ***@***.***>, Author ***@***.***> 主题：Re: [OpenImagingLab/UltraFusion] How to process the training data?(Issue #9) cecret3350 left a comment (OpenImagingLab/UltraFusion#9) Thank you for your recognition of our work. The over-exposure image is normalized to [-1, 1] to meet the requirement of VAE. We normalize the under-exposure image to [0, 1], mask with a pseudo occlusion mask and then extract the structure and color information. We also apply widely-used augmentation techniques (random crop, resize, rotate, flip) during the training process. — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.