Section: New Results

Reconstruction and gradient-based video editing

Participants : Hicham Badri [correspondant] , Hussein Yahia, Driss Aboutajdine.

Figure 14. From left to right: original image and examples of non-photoralistic rendering.

Figure 15. Top left: original image. Top right: object removal with FFT-reconstruction algorithm. Botton left: object removal with MVC (Mean Value Coordinates) algorithm. Bottom right: object removal by numerical solving of Poisson equation.

Gradient-domain methods have become a standard for many computational photography applications including object cloning, panorama stitching and non-photorealistic rendering. Integration from a vector field is required to perform gradient-domain-based applications and this operation must be fast enough for interactive editing. The most popular way to perform this integration is known as the Poisson equation and requires solving a large linear system that becomes more costly as the region of interest becomes larger. We propose to use an FFT-based solution and the framework of reconstructible systems instead of performing interactive local/global editing in the gradient domain on the CPU/GPU for both images and videos. See figures 14 , 15 .

• Related publication: [21] .