Apples to Apples comparison

[suryagutta]

I have doubts about the comparison of the results based on the attached image. In the image, we are comparing 64 images. But, in reality, we might be having 64 * 5 (assume 5 augmentations) = 320 images. I can argue that I can apply the same augmentations and dynamically label those new images based on the main image. Now, I will have 320 labeled images. Here, I am not talking about whether it's a positive or negative image pair as SymCLR uses. But, it's a straight forward image generation and labeling (no manual labeling). I think the supervised performance might increase and won't be worse.

[suryagutta]

The above question has been posted in "application (CA irrigation detection)". We will be tracking that thread.