How to create dataset with real and virtual image

I want to use Tensor Flow to classify parts using a camera.
So to create my dataset, to teach the CNN, I want to use 3D image.
I use a 3D generator to create image of all parts in many position. Taking pictures of all the parts in many position, that I want to classify is not possible because I have lot of differents parts.

My problem is when I use my CNN with real image, it did classify my parts.

What is the best way to classify real image with “virtual” image ?

I found lot of article, that explain the theory. I look for article about real exemple, and / or sample code.

Thanks for help.

To explore more in depth all the domain adaptation issues I suggest to take a look at:

Also if It Is a little bit old, if you want some reference implementations, check the sim2real challenge leaderboard:

Thanks for the links.
They will be helpfull