Hi everyone,
I’m working with a resnet50 model to make some predictions on some complex images. I have also a consistent quantity of synthetic images similar the real ones. Problem is that my model has to be trained only on synthetic images and be evaluated on real images. During training phase I get low validation accuracy (around 10%). So, my question is if there exists some techniques to convert one particular real image into a synthetic one or viceversa? thank you.
I know this is a open ended question. Here is my thoughts
First check how similar is the real images to synthetic ones either using some similarity measure. Because if it is not similar enough then the model is training on one distribution and testing on another distribution which is non IID assumption. So your accuracy will always be low.
Secondly converting real to synthetic can be done using multiple techniques out there like GAN or regular methods too. But the question want in the synthetic images is it just the pixel changes or noise …etc