Implementation of localized style transfer from the paper Localized style transfer.
Here the simplified approach is adopted - we do not use the CRF
(conditional random fields), which are used for smoothing and general improvement of
the segmentation task.
In order to make a style transfer one needs to pass a content image, obtain a mask on it via evaluation on the pretrained segmentation network, and the style image.
The current implementation uses MaskRCNN
with resnet-50
backbone and vgg19
backbone for the style transfer.
Function from detection
directory are copied from [PyTorch vision] https://github.com/pytorch/vision