BusterNet_pytorch
BusterNet_pytorch copied to clipboard
Dataset creation
Hello,
First of all, thank you for providing this re-implementation of this code in pytorch.
The synthetic dataset creation is based on MIT SUN2012 and MS COCO. But, as far as I understood, only MS COCO provides annotations with proper masks. SUN2012 provides bounding regions for the objects. So a mask for SUN2012 images would be a simple rectangular patch, which could look blocky and immediately obvious if copied. I am working only with COCO images for real and forgery iamges and leave SUN only for real images.
Was this the approach used in the paper?