Palette-based Recoloring* A fast, real-time and intuitive way to - PowerPoint PPT Presentation

Palette-based Recoloring* A fast, real-time and intuitive way to re-colorize photos for non-experts. * Based on the “ Palette-Based Photo Recoloring ” paper from Huiwen Chang, Ohad Fried, Yiming Liu, Stephen DiVerdi and Adam Finkelstein.

Objectives ● I'd just like to change the fox's fur. How to do it? ● How to preserve the brown tones on the fur? ● Do I need to be a Photoshop expert to do it?

Palette selection ● We want a good set of representing colors ● We use a deterministic weighted K-means algorithm to get them.

How we do it ● First we make a 15x15x15 bins in RGB space and histogram using colors in the picture. ● Use that as the input for the weighted deterministic K-means algorithm. ● The first means are set as the the bins with highest frequency. ● Distances are computed in Lab space.

Changing the K value K = 6 K = 3 K value can be used to localize color edition.

Color edition ● As seen, we find the initial palette by clustering. ● Now, we use it as a control for color edition. ● How to re-colorize each pixel in a natural way? Original Palette New Palette

Two-step Color Transfer Function ● First we define a L-based transfer function ● Then an ab-based one using that L-slide.

Single ab-transfer function ● In the case where there is just ONE palette color, we change every pixel “almost” int the same direction. ● Points near the border have to remain in gamut.

Single ab-transfer function construction. Points near the gamut frontier are set to stay in and towards C'

General ab-based transfer function ● In the general case, we blend all directions with proper weights. ● The closer the pixel color is to a changed palette color, the more similar it changes in that direction.

General ab-transfer function construction equations. For any color x

Implementation (too buggy for now)

Now it is possible to do this without dying in the trying.

Future work ● Add some spacial awareness to the algorithm. ● Fix the big obscure bugs the implementation has. Yes, those awful red pixels are not part of the algorithm. ● Software optimizations to make the program robust and even easier to use.