wither in the light

Wow... while their paper says they've only made some minor improvements on previous work, it looks to me like researchers at the University of Jerusalem have busted the field of image colorization wide open. They've developed an algorithmic technique for colorizing greyscale photos and video with minimal human input - the colorizer makes color annotation marks on the greyscale original, and the algorithm attempts to colorize the rest of the picture. There's still a need for an artistic eye in choosing the palette with which to annotate, but the grunt work is all handled by the process.

Processing times are so small that film could be colorized using multiple low-impact refining passes; this process could even be sped up by having the application evolve the most likely palette from previous frames. This is something like the digital equivalent of having numerous artists in a row, feeding celluloid film continuously from desk to desk, each artist adding small refinements until the result is good enough.

Their Siggraph paper is here (3.36MB PDF), and Matlab implementations are here (331KB zip).

[via waxy]