A neural network builds 3D hair models from a single photo by predicting strand density and matching it to a database of 50,000 reconstructed hairstyles. The method works automatically and lets you edit the result, but can't reconstruct hair hidden from the camera angle.
Questions this video answers
50,000 photographs and their reconstructions.
A single photograph.
A digital 3D hair model (geometry that can be used in movies/games).
Chapters
Dear fellow scholars, this is two minute papers. With a couple episodes ago, we finally set sail in the wonderful world of hair simulations, and today we shall continue our journey in this domain. But this time we're going to talk about hair modeling. So first, what is the difference between hair simulation and modeling? Well, simulation is about trying to compute the physical forces that act on hair strands that thereby showing to the user how they would move about in reality. Modeling, however, is about obtaining geometry information from a photograph. This geometry information we can use in our movies and computer games. We can also run simulations on them and see how they look on a digital character. Just think about it. The input is one photograph and the output is a digital 3D model. This sounds like a remarkably difficult problem, typically something that a human would do quite well at, but it would be too labor intensive to do so for a large number of hairstyles. Therefore, as usual, neural networks enter the fray by looking at the photograph and trying to estimate the densities and distributions of the hair strands. The predicted results are then matched with the hairstyles found in public data repositories, and the closest match is presented to the user. You can see some possible distribution classes here. Notice not only that, but the method is fully automatic, which means that unlike most previous works, it doesn't need any guidance from the user to accomplish this task. As a result, the authors created an enormous dataset with 50,000 photographs and their reconstructions that they made freely available for everyone to use. The output results are so spectacular it's almost as if we were seeing magic unfold before our eyes. The fact that we have so many hairstyles in this dataset also opens up the possibility of editing, which is always quite a treat for artists working in the industry. The main limitation is a poorer reconstruction of regions that are not visible in the input photograph, but I think that goes without saying. It is slightly ameliorated by the fact that the public repositories contain hairstyles that make sense, so we can expect results of reasonable quality even for the regions we haven't seen in the input photograph. As always, please let me know below in the comments section whether you have found everything understandable in this episode. Was it easy to digest? Was there something that was not easy to follow? Your feedback, as always, is greatly appreciated. Thanks for watching and for your generous support and I'll see you next time. Ra.
