Gradient-Domain Fusion

The goal of this project is to seamlessly blend an object from one image to another using the Poisson Blending method described in the Perez et al. 2003 paper.

Poisson Blending is a guided interpolation approach that re-creates the target region given boundary constraints under guidance of the gradient field of the source region.

Poisson Blending

(Original flickrMolly Goossens / CC BY NC)

Comparison with cut-and-paste

My friend Brian created this image below using the cut-and-paste method (and possibly some feathering). Can we use poisson blending to improve this?

Target image:

Source image:

Another example

ACM office in Siebel Center

Failure case

One reason for this result is because my implementation doesn’t set boundary constraints at edge locations. Another reason is because only the gradients of the original image is preserved. So when the surrounding area has a significantly different light intensity, interpolation will distort the source image.

Mixed gradients

(Photo by flickrDavide Restivo / CC BY SA)

Left: The mixed gradients method preserves transparency.

Failure case

Bells & Whistles


Laplacian Pyramid Blending

Classic example. From level 1 to 7.

vs. poisson blending:

Compared to the poisson blending example, the result from laplacian blending has no color distortion, however the pixels away from the boundary did not blend.