- Artur Kasymov
- Daniel Dobrowolski
- Ewelina Jamrozik
- Przemysław Przybysz
We want to use Proptypes to generate / modify images. We focused on simple images from the AO-CLEVR dataset.
AO-CLEVr is a synthetic-images dataset created by NVIDIA. Images are created to "easy" Attribute-Object categrosies, based on the CLEVr framwork (Johnson et al. CVPR 2017).
Each object has several attributes:
- Shape [sphere, cube, cylinder]
- Color [red, purple, yellow, blue, green, cyan, gray, brown]
- Size [small, large]
- Material [rubber, metallic]
- Random position
- Random light
Each Shape X Color pair consists of ~7,500 images.
Samples:
