Deep-Photo-Style-Transfer

The project intends to implement the original paper on Deep Photo Style Transfer by Luan. et. al. We have built a pytorch implementation of the original model built by the authors.

1. We have modified and built upon the publicly available pytorch implementation of Neural Style Transfer by Gatys. Et. al

Neural Style Transfer

2. We have used the torch implementation of Matting Laplacian available on github.

Matting Laplacian

Our model makes the following assumptions:

1. Pretrained VGG19 as feature extractor for the images
2. Conv4_2 used for content loss and Conv1_1, Conv2_1, Conv3_1, Conv4_1, Conv5_1 used for style losses.
3. The input image is already photorealistic.
4. Weights as mentioned in the paper.

Results obtained are attached in the powerpoint presentation attached in this repository.

Powerpoint Presentation

References

Original Paper:

Luan, Fujun, et al. "Deep photo style transfer." Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2017.

Open Source Resources:

https://pytorch.org/tutorials/advanced/neural_style_tutorial.html

https://github.com/MarcoForte/closed-form-matting

https://github.com/LouieYang/deep-photo-styletransfer-tf

Credits

Rohit Jain

Vaishnavi Kotturu

Khushi Gupta