Research On The 3D Garment Deformation Based On Motion Type

3D garment deformation and animation synthesis is a vital technique in computer graphics,which is widely used in video,games,virtual try-on and etc.With the growing demand in these fields,how to fastly generate diverse clothed 3D human animations is becoming the hotspot research.Traditional garment authoring workflows mostly construct a physical equation of the cloth to solve the dynamics status of the cloth at different times.Hence,the clothing deformation pattern driven by human motions cannot be applied to the new scenes,with the result that the animators have to set different simulation environments repetitively and adjust the visual effects iteratively,which is laborious,time-consuming,and computationally expensive.On the other hand,established data-driven methods generally formulate the authoring of garment animation as a translation task from human motion to clothing animation,which means that the synthesized garment deformation effects mostly depend on the given human motion.Therefore,when generating multi-style dressed human animations utilizing these methods,the animators need to pick up the ideal human motions first,and then synthesize the corresponding garment deformation.Such methods lack generative capacity and are not capable of generating diverse clothed human animations according to the given high-dimensional semantic parameters.To that end,we conduct the research on 3D garment deformation and animation synthesis based on human motion type from two aspects:the correlation between the human motion pose with different semantic motion labels and the garment deformation,and the generation of diverse clothed human motions driven by semantic motion labels.Specifically,we respectively learn to model the probability distributions of human pose and motion sequence under different motion semantics,and build the mapping between the distributional points and garment deformation.By specifying the human pose with motion semantic or the semantic motion label,the proposed workflow is able to generate diverse dressed human animations.The main work of this paper is as follows:(1)We study a method for 3D garment deformation prediction driven by body pose.Different from the existing methods that directly learn the mapping between the human motion and the garment deformation,we hypothesize that the body poses with different motion semantics and the clothing deformation approximately complies with the same probability distribution model.We firstly construct a body pose encoding network to extract the semantic features from the body poses,and further to fit the posterior distributional model from amounts of garment deformation instances.Then we build the mapping between the sample point of the distributional model and the garment deformation utilizing a garment decoding network.By giving the body pose parameters,our method can synthesize the corresponding garment deformation.Besides,a few more constraints are imposed to modify the posterior probability distribution model to a conditional posterior probability distribution model,which is able to synthesize various visual garment deformation effects.(2)We study a multi-style dressed human animation generation method driven by semantic motion type.We combine a temporal network with a generative model,which is capable of generating diverse clothed human animations according to the given semantic motion label.Inspired by the relationship between human motion and garment animation in real scene,we regard garment animation as the ramification of human motion,so we disentangle human animation generation and clothing animation synthesis.For human motion generation,we combine a Transformer-based network with the conditional variational auto-encoder to learn the motion semantic features of different kinds of human motions and fit a latent space encoding the posterior probability distribution of the semantic features,which is able to generate multi-style human motions by specifying the semantic motion label.For garment animation synthesis,we take it as a pure translation task from human motion to garment animation,and learn the mapping utilizing a Transformer-based network that adopts the self-attention mechanism to ensure the temporal consistency of the garment deformation sequence.Since the two branches are trained separately,there may exist interpenetration between the generated motion sequence and the synthesized garment deformation sequence.To alleviate this phenomenon,we further introduce a learnable BOS(beginning of sequence)strategy and a self-supervised collision handler,which is important to generate a sequence of dressed human animation with realism and temporal coherence.(3)We build a garment animation synthesis system.To meet the demands of the animators,such as visualization and interactivity,etc.,we adopt a three-tier architecture to design and implement the system.It is composed of a data presentation layer,a business logic layer,and a UI layer,providing a lot of authoring functions,such as garment deformation prediction,animation synthesis,rendering management,animation management,and so on.