상세정보

Video object removal, when given a video, aims to delete a particular object from the image and to plausibly fill the background in place. Image object removal has been studied tremendously since the rise of deep learning, whereas research in the video domain has not been studied due to challenging points of preserving spatio-temporal coherence and a large amount of computation. In spite of these difficulties, most media in real life are composed of video, so research is necessary.
In this work, we propose a network that performs video object removal using edge and optical flow information based on GAN network. The proposed network consists of three sub-networks; learning the edge information to create a background contour that is likely to be at the location of the object to be erased, translating the background pixel at the position of the object through the temporal relationship of the frames in the input video, and inpainting an image using edge information in a reduced empty space in preceding steps. 
Our method is evaluated on DAVIS datasets qualitatively and quantitatively, achieving the state-of-the-art performance in terms of object removal quality and speed. 

Contents
1.  Introduction	- 1 -
2.  Related Works	- 5 -
2.1  Video Inpainting	- 6 -
2.2  Edge with Inpainting	- 12 -
2.3  Optical Flow	- 15 -
3.  The Proposed Method	- 17 -
3.1  Edge Generator	- 20 -
3.2  Flow Guided Frame Warping	- 23 -
3.3  Video Completion Network	- 25 -
4.  Experiments	- 28 -
4.1.  Implementation Details.	- 28 -
4.2.  The Quantitative evaluation with the State-of-the-art Methods	- 29 -
4.3.  The Qualitative evaluation with the State-of-the-art Methods	- 31 -
4.4.  User study	- 38 -
5.  Conclusion & Future Works	- 39 -
References	- 40 -