Research On The Data Driven Based3D Virtual Try-on Techniques

3D virtual try-on technology has been widely used in the domains of e-business, GCAD, movie and3D computer games, etc. The commonly-used physics-based method is so complicated in computation that lowers the efficiency of simulation, which has restricted its development and application. Therefore, research on a more efficient and real3D virtual try-on technology plays an important role in promotion of relative industries.Based on current research of3D virtual try-on, this paper presents a data-driven based3D virtual try-on scheme. By analyzing the preexisting cloth-human sample data, this method builds the mapping relationships, usually in the form of simple matrixes arithmetic, between the cloth deformation and the human dynamics (usually including human posture, human motion speed, etc.). The final results are obtained after the simple mapping calculation.A physics-geometry mixed method for obtaining the dressed human dynamic data is proposed. The cloth model is redefined into the tight, the float and the loose model by extracting the cloth region segmentation factor which is composed of the dynamic cloth clearance and the local detail complexity. The cloth tight model and the float model are deformed by adopting the skin method and the as-rigid-as-possible method respectively. The loose model, on the other hand, is using the physics-based method for achieving more wrinkle details. A parameter is set to dynamicly adjust the proportion of each model. The sample data is obtained by finally applying different collision handling onto their corresponding models.A data mapping model based on the human dynamics and the cloth history deformation status is proposed. Based on the observation that the triangles of cloth local patches always undergo similar deformation regularities, both the cloth and the human model are segmented into several independent patches. The cloth simulation is completed by synthesizing the patch-form deformation and the triangles wrinkles supplementation. The linear and the non-linear regression methods are adopted to obtain two mappings:the mapping between the cloth patches’deformation and the human posture; the mappings between the triangles wrinkles supplementation and the human speed as well as acceleration. Finally the cloth history deformation is utilized to filter the cloth deformation in time axis.A collision handling method based on the model collision cell is proposed. Based on the observation that the triangles of cloth local regions intersect with the similar regions of the human model, the models are redefined into several collision cells by using the hierarchical clustering method. Each collision cell is organized into an independent BVTT (bounding volume test tree), and every two cells with intersection are matched as a cell pair. The probability whether a cell would intersect with the other is extracted by adopting the Naive Bayes method. A static-dynamic mixed task distribution mechanism is proposed to make the computation load averagely distributed onto each thread and the detection process parallelly implemented. During the real-time simulation, only the cell pairs with high intersection probability are detected.Based on the research, the visual try-on platform is developed and the cloth dynamical simulation is instantiated. Finally the contents are summarized and the future works are presented based on the drawbacks of this paper.