Image Compositing and Editing - Computer Vision for Visual Effects - page 76

(a) (b)

(c) (d) (e)

Figure 3.19. A failure case of PDE-based inpainting. (a) The original image. (b) The inpainting
mask. (c) The result after 20,000 iterations of PDE-based inpainting. (d) The result after 200,000
iterations of PDE-based inpainting. The inpainting region is unacceptably blurry. (e) Poisson
compositing with a guidance vector field of 0 inside the inpainting mask, giving a similar result.

inside . This would result in a second-order PDE (i.e., a Laplace equation with

Dirichlet boundary conditions) as opposed to the third-order PDEof Equation ( 3.30 ) . 6
PDE-based inpainting techniques are a reasonable choice for certain visual effects
scenarios, such as painting out thin wires holding up a stunt performer. However,
a major drawback of PDE-based techniques is that the interior of the inpainted
region is inevitably smoother and blurrier than its surroundings, leading to unac-
ceptable visual artifacts when a large hole is located inside a textured region, as
illustrated in Figure 3.19 . The patch-based methods we discuss next do not have this
shortcoming.

6 Since inpainting can take many iterations to converge, using the Poisson approach is also likely to
be much faster.