Implementation of 2D-to-3D Prediction Network for Point Cloud Clothing Alignment

The virtual try-on of garment requires to capture the outfit of the three-dimensional human body with the true measurement. Point-based modeling can estimate the geometry of diverse clothing outfits in point clouds with the modelingof human body shapes. In this paper, a 2D-to-3Dhuman pose prediction network is used to estimate 3D human shapes. To account for the absence of scale parameters, a scale optimization module is introduced to regress the scale value using the vertices generated by I2L-MeshNet and point-based modeling. Additionally, a pelvis center alignment step is incorporated to ensure the human body is accurately aligned with the center of the image. A dataset with human models in T-pose and random pose are collected for evaluation. The appearance reconstructionresults are then compared with manually cropped images and auto-center positioning images using YOLOv8. Measurement metrics such as Intersection over Union and ratio of overlap are used to identify the effectiveness of three different clothing alignment techniques. This paper demonstrates the proposed pipeline can accurately aligns the outfit to the target human body and deliver the best appearance reconstruction result.