Colorization Chapter
The portrait experiment
Second Summary
Second set of experiments
More parameters are varying : the cost function of spatial changes, and the degree of smoothing of the estimated color probabilities in the color space.
The reason for this second exploration of the parameter space is that, in the previous set of experiments, the boundaries of the color regions are better when high spatial coherency is required, but the strange counterpart was a color averaging, or fading, to grayer colors. The aim here was to keep vivid colors even when spatial coherency is enforced.
- Apparently, spatial variations are just a bit better if one enforces them with a cost varying faster as a function of the estimated color gradient norm.
- And apparently, colors are much better if one smoothes more the estimated color distributions in the color space, while I thought initially they would be better with less smoothing.
Caption :
- + : more color smoothing (the same amount for all pictures tagged with +)
- * : same amount of smoothing as in previous experiments
- - : less color smoothing
- s : more constrast in the spatial coherency cost (black is black, white is white, grey is whiter)
- number : spatial coherency weight / 1 000 000 (the value depends on the cost function, so should not be compared with previous experiments without enforced contrast)
+ s 0.7 |
+ s 1 |
+ s 1.2 |
+ s 1.5 |
+ s 2 |
* s 2 |
- s 0.5 |
- s 1 |
- s 1.5 |
- s 2 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
+ s 0.7 |
+ s 1 |
+ s 1.2 |
+ s 1.5 |
+ s 2 |
* s 2 |
- s 0.5 |
- s 1 |
- s 1.5 |
- s 2 |
Pick your favourite value of the spatial coherency parameter !
Remember the edges predicted (they don't depend on the parameters above, but the cost function that uses them can vary) and the original images:
Back to the colorization chapter page
Back to the main page