search

chenhsuanlin / spatial-transformer-GAN

ST-GAN: Spatial Transformer Generative Adversarial Networks for Image Compositing :eyeglasses: (CVPR 2018)
MIT License
334 stars 72 forks source link

3D cube synthesized daataset #21

Closed ipod825 closed 5 years ago

ipod825 commented 5 years ago

Could you share the code on how you create the 3D cube dataset? It would be very helpful to test the model without running on a large dataset. Thanks.

chenhsuanlin commented 5 years ago

Unfortunately I no longer have access to the code to generate the 3D cube dataset, but I've uploaded the rendered training set and test set. Hope this helps!

ipod825 commented 5 years ago

Thanks! This helps a lot. May I ask what (perhaps python?) package did you use to generate these data?

ipod825 commented 5 years ago

The training set does not contain "real" images, only but "fg" and "bg" images.

chenhsuanlin commented 5 years ago

Corresponding foreground/background images were rendered from the same viewpoint and configuration, so directly compositing the two together gives you the "real" images. Fake images were generated by adding geometric perturbations on the fly (the same as in the released code).

I generated this dataset using Matlab. Basically it's just drawing colored 3D planes to form the 3D cube/room and rendering from different viewpoints, so it could also be done easily in Python.