We use PASCAL VOC dataset for the foreground.
The data is organised as follows:
├── background
└── VOC2010
├── Annotations
├── ImageSets
│ ├── Action
│ ├── Layout
│ ├── Main
│ └── Segmentation
├── JPEGImages
├── SegmentationClass
└── SegmentationObject
Run this to generate foreground.
python generate_foreground.py
We crop DIV2K dataset to 256*256 for background images.
Run this to generate RGB datasets (./trainset0908).
python generate_motion.py --config ./config.yml
-
step: the original images with sub-pixel motions
-
gt: the ground-truth of the alpha times maginified images
-
The number behind underscore is magnification factor.