It reminds me of what google did to improve their search for images: https://crowdsource.google.com/imagelabeler/category

The things you draw here is new input for their model.