关键词: 三维服装提取, 三维点云, PointNet++, 点云分割, 边界优化

Abstract: Garment extraction is the process of accurately extracting clothing information from 2D images or 3D models, aiming to achieve digital storage, search and outfit matching of clothing. With the rapid development of digital technologies, garment extraction holds broad application prospects in fields such as virtual try-on, fashion recommendation and intelligent clothing design. However, when 3D point cloud scene segmentation methods are applied to garment extraction, the boundary regions between the human body and clothing, as well as between different parts of the clothing, exhibit complex geometric details and variations. This results in poor boundary segmentation performance, subsequently affecting the overall segmentation accuracy. To address this issue and achieve more accurate 3D garment extraction and segmentation, this paper proposes a 3D garment point cloud segmentation scheme based on point cloud segmentation and boundary optimization, leveraging an improved PointNet++ method. Firstly, this study constructs a high-quality 3D dressed human point cloud model dataset. Based on the SynBody synthetic human dataset, a total of 583 virtual dressed human samples are selected, covering five common clothing types such as tops, pants and skirts. To ensure the accuracy and diversity of the dataset, each sample is manually segmented and labeled using a multi-level segmentation strategy, with explicit annotations for three human body parts such as head, torso and limbs and five clothing types. The construction of this dataset provides effective training data for deep learning, laying the foundation for subsequent algorithm optimization. In terms of the algorithm, this study proposes a step-by-step training and joint prediction method based on the PointNet++ model to significantly improve the segmentation performance in boundary regions, thereby enhancing the overall segmentation effect. Specifically, a boundary identification method relying on K-nearest neighbor (Knn) distance calculation is first introduced. By computing the neighborhood distances of each point in the point cloud, boundary points within the coarse segmentation results are identified. Subsequently, step-by-step training and joint prediction are carried out separately for the overall model of the dressed human body and the boundary regions. During the training process, the segmentation results of the overall model serve as the initial input, and the optimization of boundary regions is further refined through joint prediction. This method not only effectively captures the complex geometric features of boundary regions but also avoids the misclassification issues that commonly occur at boundaries in traditional methods. To validate the effectiveness of the proposed method, this paper conducts boundary optimization experiments and overall segmentation performance evaluations. The experimental results show that compared to baseline method, the proposed method achieves a 32.74% improvement in Overall Accuracy (OA) and a 34.25% improvement in mean Intersection-over-Union (mIoU) in boundary regions. Additionally, the OA and mIoU for the overall segmentation are enhanced by 1.21% and 0.89% respectively. These results fully substantiate the improvements in boundary region segmentation and overall segmentation accuracy achieved by the proposed method.

Key words: 3D garment extraction, 3D point cloud, PointNet++, point cloud segmentation, boundary optimization