Automatic hair modeling from one image

Hair is one of the most critical characteristics of a person in the process of digitizing characters, but at the same time, hair modeling is still a very challenging task due to the diversity of hairstyles and the overlap among hair. We introduce a method to automatically generate 3D hair geometry from a front hair image, which can recover the outline and details of the hair geometry. We designed a encoder-decoder convolutional neural network which takes the 2D orientation field from a hair image as input, and output the characteristic hair geometry. Then we use the characteristic strands to search for eligible hairstyle from Hairstyle-database, and fuse retrieved hair model to get the nal hairstyle. This pipeline can automatically recover hair geometry from a front hair image without any supplementary information. Experimental results show that our approach achieves realistic reconstruction effect from real Internet pictures and self-portraits.