Research On Generating Catwalk Motions For Virtual 3D Fashion Shows

As a new way of showcasing fashion in the "post-internet era",3D virtual fashion shows have become an important means of fashion display and marketing,especially with the rise of the metaverse and the global spread of the COVID-19 pandemic.Current implementations of virtual fashion shows mainly rely on special effects production and animation simulation.However,special effects production is costly,and animation simulation cannot meet the demand for diverse model runway motions.Combining 3D human pose estimation and motion generation technology,learning model runway motions and styles from runway videos and automatically generating diverse motions provides a new approach for implementing virtual fashion shows.However,during the model’s runway process,clothing occlusion and self-occlusion of joints are common,making it difficult to predict the position of occluded joints.Moreover,the loss of some joints due to their complex interrelationships can also affect the prediction of unoccluded joints,ultimately affecting the quality of motion generation.Additionally,different styles of clothing require matching runway motions by models to interpret the meaning of the clothing.This article focuses on the research of model motion generation for 3D virtual fashion shows to address the above issues.A two-stage 3D model pose estimation method based on high-resolution representation learning and graph convolution is proposed to address the issues of clothing occlusion and selfocclusion of model joints in fashion runway scenes.For the problem of insufficient feature extraction and occlusion in 2D model pose estimation,a high-resolution 2D model pose estimation network based on multiscale attention is proposed.For the problems of oversmoothing after network stacking in 3D model pose regression and the inability to extract local semantic information of joints in 3D pose regression,a graph convolution 3D pose regression method based on residual connections and attention mechanism is proposed.Experimental results show that the 3D model pose estimation method based on high-resolution representation learning and graph convolution has a testing error of 56.6mm on standard datasets,which outperforms the current mainstream 3D human pose estimation methods on the standard test set and provides a good data basis for subsequent motion generation.We propose a model for generating model motions for different clothing styles based on unpaired style transfer.To address the processing and matching consistency of style and content features for different style runway motions,we design a dual-branch feature processing generator,which performs channel-level feature processing on style and content features,and uses Adaptive Instance Normalization algorithm to match consistency between style and content features to reduce losses in feature fusion.We design a loss function consisting of content consistency loss,adversarial loss,and regularization loss to improve the quality of motion generation during the mutual adversarial process between generator and discriminator.Experimental results show that the unpaired style transfer-based model for generating model motions achieves a similarity result of 0.87 on the training set and 0.63 on the test set,as evaluated by cosine similarity,both of which are higher than the results of the current similar methods,effectively preserving the content and style features of the input runway motions.Based on the above research,a prototype system for generating model motions in 3D virtual fashion shows was designed and developed.First,the actual needs of the system were analyzed,and then the architecture of the prototype system was designed and the functions of each module were implemented.Finally,application verification was carried out through engineering examples,providing a platform and tools for generating model motions in 3D virtual fashion shows.