Research On Garment Cloth Simulation Based On Adaptive Character Feature Network

In real-time applications such as computer simulation,video games,and digital art,there are often images of character garment wrinkling along with the characters' movements,or flying up and down due to gusts of wind.As one of the most common phenomenon in daily life,real cloth animation can significantly enhance the realism of the virtual world.In the past few decades,many scholars have proposed many solutions using traditional physical and non-physical methods.In recent years,methods based on neural networks have emerged.The current method is computationally expensive,and in most real-time applications at present,cloth simulation can only be partially applied or approximated by other methods.The research content of this paper is cloth simulation based on neural network method under the condition of real-time operation,and the details are as follows:Firstly,proposed a real-time garment cloth simulation method based on feedforward neural network.Given a character and garment,the GAFNN(Garment Animation Feedforward Neural Network)model consisting of a garment feature extraction network and an animation inference network is designed.The skeleton animation data sequence of the character at the current and past moments is input,and the vertex animation of the garment at the current moment is output.Because of the lack of relevant public data sets,the data set Chara-Garment Motion was designed and produced.The feasibility of the simulation method is verified through experiments: a visually credible simulation effect is achieved on the basis of meeting the performance requirements of real-time operation.The single-frame simulation time consumption is less than 2% of the traditional method,and the vertex error of the inference result is within 5cm.Secondly,proposed a real-time garment cloth simulation method based on variational recurrent neural network.Aiming at the problem of vertex animation jitter in GAFNN,the M-VRNN(Motion Variational Recurrent Neural Network)model is proposed,which inputs the current character animation and the garment features of the previous frame,and outputs the current garment vertex animation.Compared with GAFNN only considers the temporal coherence of the input,this method further considers the temporal coherence of the output.M-VRNN is an improvement of the variational recurrent neural network model,adding character animation data input when performing a priori inference.This method does not predict the value but the probability distribution of the predicted value,and has more posterior information,which enhances the robustness of the model.Experimental results show that this method solves the jitter problem and achieves better visual effects.