We examine the sensitivity of our algorithm to the initialization of the color mean ν_k. Depending on its initial state the algorithm reaches slightly different local minima and the number of iterations varies.

In the video linked below we examine four different initializations:

• Gray: initialize ν_k to middle gray
• Bicubic: initialize ν_k to a bicubic rescaling of the input
• Zero: initialize ν_k to zero (shows as blue due to RGB ↔ CIELAB conversion)
• Random: initialize ν_k and Σ_k to random values.

We use the "gray" initialization, which worked best in our tests.

Click here to open video
(see "analysis_convergence" directory if opening in the browser fails)