A lot of synthetic document data is still just simple text overlays. It misses real physical effects — folds, stretch, paper deformation, natural lighting, shadows etc.
I work on photorealistic image synthesis for documents. More realistic training images make models noticeably more stable in real-world conditions.
Curious if NanoNets has considered training/releasing models on higher-realism synthetic data like this?
https://www.alrowilde.com/
A lot of synthetic document data is still just simple text overlays. It misses real physical effects — folds, stretch, paper deformation, natural lighting, shadows etc.
I work on photorealistic image synthesis for documents. More realistic training images make models noticeably more stable in real-world conditions.
Curious if NanoNets has considered training/releasing models on higher-realism synthetic data like this?
https://www.alrowilde.com/