On this page we compare our algorithm against
Gerstner et al., Pixelated Image Abstraction, NPAR 2012.
Our algorithm was not designed for palette reduced results on natural images. We achieve palette reduction in a post process using K-means clustering in CIELAB color space. The K mean vectors are initialized as in this pseudocode:
Gerstner et al., Pixelated Image Abstraction, NPAR 2012.
Our algorithm was not designed for palette reduced results on natural images. We achieve palette reduction in a post process using K-means clustering in CIELAB color space. The K mean vectors are initialized as in this pseudocode:
mean[0] = random pixel color
for i = 1 to K-1
mean[i] = most dissimilar color from mean[0, ..., i-1]
end for
for i = 1 to K-1
mean[i] = most dissimilar color from mean[0, ..., i-1]
end for
K-means is run with this initialization until convergence.
Astronaut
 Input |
 Gerstner's result (16 colors) |
 Our result (16 colors) |
 Our result (not quantized) |
Dog
 Input |
 Gerstner's result (16 colors) |
 Our result (16 colors) |
 Our result (not quantized) |
Giraffe
 Input |
 Gerstner's result (16 colors) |
 Our result (16 colors) |
 Our result (not quantized) |
Indian
 Input |
 Gerstner's result (16 colors) |
 Our result (16 colors) |
 Our result (not quantized) |
Lenna
 Input |
 Gerstner's result (16 colors) |
 Our result (16 colors) |
 Our result (not quantized) |
Man
 Input |
 Gerstner's result (16 colors) |
 Our result (16 colors) |
 Our result (not quantized) |
Motorcyclist
 Input |
 Gerstner's result (16 colors) |
 Our result (16 colors) |
 Our result (not quantized) |
Obama
 Input |
 Gerstner's result (16 colors) |
 Our result (16 colors) |
 Our result (not quantized) |
Tower
 Input |
 Gerstner's result (16 colors) |
 Our result (16 colors) |
 Our result (not quantized) |
Back to previous page