There has been previous work on automatic colorization of images, a fairly impressive process that takes a black and white image and converts into a colour one. This uses DNN’s and their ability to guess at probable outcomes based on massive amounts of trained data. It’s pretty cool, but it can sometimes go wrong — water is blue because it usually is, so it’s a good guess. Peppers are… mostly orange?
Well, a paper and talk at Berkeley (now surfacing due to SIGGRAPH) highlights guided colorization; that is, a user supplies the DNN with some initial information on what colour an item should be, and the network fills in the rest. This provides impressive results, as expected, but more impressive is the YouTube video that shows the researchers’ software and its ability to on-the-fly change the colours based on user input.
This is an excellent demonstration of the better capabilities of DNNs that are available today; not necessarily self-serving, autonomous assistants, but very intelligent tools that can massively enhance productivity of the human operators.
