Tensorflow implement for DeepWarp: Photorealistic Image Resynthesis for Gaze Manipulation
This is a re-implement of paper DeepWarp by Tensorflow. The results of my implement is slightly worse than the original paper which you can find it in DeepWarp Demo Page. Actually in general, I achieve the basic function for moving gaze. Some of the results and its drawbacks will be shown in behind.
But in the another hand, there are some differences between my implement and the paper.
self.loss = self.coarse_loss + self.output_loss
For a better experience, I used some mess codes implementing a GUI to visual the results. All of these can be found in gui.py
2018.06.26, the first release version of DeepWarp, it finish the basic function for moving gaze.
The first one is the original image. The other two are the generated images. In the GUI, set the output path and click buttons vertical and horizontal to generate the two. As you and see, the eye white is not very good. And also, when the value reach the boundary(vertical is [-30,30], horizontal is [-60, 60]) the results become unreasonable.