HAO-CNN:Prediction Of Hair Volumetric Vector Field From A Single Image Based On Adaptive O-CNN

Hair reconstruction is one of the long-term research topics in the field of computer graphics.It is an important technique for applications such as virtual reality,virtual fitting,and animation character modeling.Unfortunately,high-quality hair reconstruction is still a challenging task due to the complexity of hair geometries and the diversity of hair styles.Because of the convenience of acquiring photos of all kinds of hair styles,image-based hair reconstruction becomes one of mainstreams of hair reconstruction.Among this kind of methods,representing hair models as volumetric vector fields is particularly suitable to convolutional neural networks and appears promising in achieving good reconstruction results.However,performing CNN on high-resolution volumetric vector fields is both time and memory consuming.In this paper,we propose a so-called HAO-CNN,a deep neural network to reconstruct the volumetric vector field from a single image.Specifically,given a single-view image as input,our algorithm first segment the hair region,and further extract the hair orientation map as well as body torso map as the input of HAO-CNN.Then HAO-CNN is responsible for mapping the input image into hairstyle space,and output the volume vector field of the target hairstyle in form of adaptive octree.Finally a improved hair growth algorithm restore the hair strands from volumetric vector fields to obtain hairstyle result.Experimental results show that the proposed method input and generate a realistic hair model that matches the input hairstyle in the term of appearance and outline.Compared with other volumetric vector field based methods,our method is more efficient in both time and memory consumption.