Abstract by junseong Ahn
Inpainting Edge of High Resolution Images under control
Recent advances in deep learning have yielded significant improvements in filling large holes with contextual aware details. Even though many of these techniques can generate reasonably smoothed structures, they have three limitations: 1. they can't control the result. 2. they can't fill in the edges part 3. they fail for higher resolution images.
In this presentation, we propose the two-stages adversarial manipulable model composed of an inpainting model with a multi-scale synthesis approach with texture constraints. Our model can resolve three limitations that recent techniques have. We evaluate our method end-to-end over MPII human pose dataset. We show our approach produces more coherent results than other techniques, especially for edge inpainting and high-resolution images.