image-inpaintingcolab-notebookhigh-resolutioncolabgenerative-adversarial-networkscnngenerative-adversarial-networkganfourier-transformfourier-convolutionspytorchfourierinpainting-methodsdeep-neural-networksinpainting-algorithmdeep-learninginpaintingcomputer-vision
You can not select more than 25 topics
Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
Roman Suvorov
552cd55c94
|
3 years ago | |
---|---|---|
.. | ||
countless | 3 years ago | |
README.md | 3 years ago | |
__init__.py | 3 years ago | |
mask.py | 3 years ago |
README.md
Current algorithm
Choice of mask objects
For identification of the objects which are suitable for mask obtaining, panoptic segmentation model from detectron2 trained on COCO. Categories of the detected instances belong either to "stuff" or "things" types. We consider that instances of objects should have category belong to "things". Besides, we set upper bound on area which is taken by the object — we consider that too big area indicates either of the instance being a background or a main object which should not be removed.
Choice of position for mask
We consider that input image has size 2^n x 2^m. We downsample it using COUNTLESS algorithm so the width is equal to 64 = 2^8 = 2^{downsample_levels}.
Augmentation
There are several parameters for augmentation:
- Scaling factor. We limit scaling to the case when a mask after scaling with pivot point in its center fits inside the image completely.