A question about loading auxiliary images

[EdyQiu]

Hi, I want to implement an idea that when loading each image in the dataset, its edge-extracted version is also loaded.
If I simply change the annotation file, it seems random sampling would stop me from achieving the goal.
I also tried to modify the train-pipeline, but still feel hard to find a good way to achieve my goal.

Any help about how could I modify the code would be very helpful!

[Ezra-Yu]

its edge-extracted version is also loaded.
Do you mean using the edge-extracted to do classification?

1. Use YourDataset without LoadImageFromFile transforms refer to the tutorial. return {'img': img, 'gt_label': gt_label} as https://github.com/open-mmlab/mmclassification/blob/fb16bdc6a25f96843f8fb97b02e45872c61dc566/mmcls/datasets/cifar.py#L91-L94
2. Add a new Transform to get the edge-extracted version.

[EdyQiu]

Thanks for your reply!

The edge-extracted version is not used for classification directly, I want to use it to enhance the edge information of the original image, and use the mixed image for classification. (Which means the label of the edge-extracted version does not matter)

I think you are suggesting me to write a new transformation to replace "LoadImageFromFile" transformation in the training pipeline right? meanwhile, I also need to write a new dataset class, or modifying the "load annotation" function of the original dataset class.

[Ezra-Yu]

Yes, If your dataset is very small as minist and cifra, I think you can just implement a new Dataset class to replace the CustomDataset and 'LoadImageFromFile'.
If your dataset is very big. modify the CustomDataset to save the edge-extracted images info and add a "LoadMixedImageFromFile" transformation.