Research On Fast Fur Generation

Fur widely exists in various films and games,and it plays a vital role in shaping characters and building realistic scenes.Due to the complex geometry and wide shape variations of fur,generating realistic fur has become one of the hot and challenging topics in computer graphics.Fur generation research mainly focuses on human hair,and few works focus on the fur attached to the surface of fur objects.However,such objects play a vital role in building realistic scenes.With the development of deep learning,some methods apply deep learning to fur generation.Such methods are highly data-dependent,but few available open-source datasets currently exist.Given the above problems,this paper focuses on the fast fur generation,taking human hair and the fur attached to the surface of fur objects as research objects.A geometric method and a data-driven method for fast fur generation are respectively proposed,and corresponding fast fur generation systems are implemented.The main work and innovation points are as follows:1.For human hair,a fast fur generation method based on geometry processing is proposed,which can generate enough hair data to solve the problem of data shortage.Some works apply deep learning to human hair generation.However,such methods are highly data-dependent,and few open-source datasets are known.In order to solve the problem of data shortage,this method generates new hair data with a natural appearance according to the reference hair model and editing parameters provided by the user.Generated hair data is aligned with the widely used human parametric model SMPL.This method can generate enough data for deep learn-ing applications such as hair modeling and human modeling,and promote the development of related fields.2.For the fur attached to the surface of the fur object,a data-driven method is proposed to solve the problem of the lack of research objects in current fur generation research.This method estimates fur parameters based on images to help artists design and create better.It starts from a real application scenario and uses fur parameters in the rendering plug-in to represent the fur.The process takes an image containing fur objects as input and uses a deep learning model to esti-mate its corresponding parameters quickly.Related datasets are created to explore the influence of different factors(such as fur color,camera view,fur carrier topology)on the fur parameter estimation.Data augmentation,view-invariant features,and pre-trained model initialization are introduced to improve the method’s performance.The effectiveness of the proposed method is verified from the perspective of quantitative and qualitative evaluation.Experiments show that this data-driven method can effectively estimate fur parameters corresponding to the fur object in the image.3.Corresponding fast fur generation systems are implemented based on the two methods above,which can quickly generate corresponding fur data in seconds and milliseconds.The fast fur generation system based on geometry processing can generate new hair data according to the reference hair model and editing parameters.Hair data can be generated in seconds using pre-computation,parallel method,and efficient data structure.The data-driven fast fur generation system bridges the domain gap between images by pre-processing image module and applies prior knowledge to small sample data with pre-trained model initialization.The deep learning model estimates the corresponding parameters in milliseconds,helping artists design and create better.