基于点云迁移的人体点云位置及颜色补全

Human Body Point Clouds Location and Color Complement Based on Point Clouds Migration

由于三维扫描设备获取的点云存在目标遮挡、设备视角等问题,致使点云有所缺失。点云补全是为下游任务进行点云预处理的重要工作。然而现有的点云补全工作集中在对简单物体的补全上,缺乏对较复杂人体点云补全的研究。为将点云补全任务引入下游人体重建相关任务中,本文提出了一种由粗到精的人体点云预测方法,对人体位置及颜色缺失信息补全。通过创建具有精细服饰纹理的人体点云数据集,采用点云迁移的网络结构,充分利用点云全局粗糙信息,结合多层网络预测完整点云,优化了上采样方案;然后,在上采样后的点云上进行了颜色补全,通过部分点云颜色信息进行特征值扩散和提取;最后,通过一个多层感知机进行颜色预测。在人体数据集上的实验结果表明,该方法相比现有主流方法不仅在客观指标上具有更好的表现,主观质量上也能保证获得更加完整、边缘清晰的点云。

The point clouds obtained by 3D scanning equipment have problems,such as target occlusion,capture perspective,etc.,resulting in the lack of point clouds.Point cloud completion is a significant work for downstream tasks.However,the existing point cloud completion work focuses on the completion of simple objects,and there is a lack of research on point cloud completion of complex human body.In order to introduce the point cloud completion task into the downstream human reconstruction,this paper proposed a coarse to fine human point cloud prediction method to reconstruction human point cloud position and color.The method in this paper created a human point cloud dataset with fine clothing texture,adopted the network structure of point cloud migration,made full use of the global rough information of point cloud,and combined multi-layer network to predict the complete point cloud,so as to optimize the up-sampling scheme;Then,the color was completed on the point cloud after up-sampling,and the eigenvalues were diffused and extracted through part of the point cloud color information;Finally,a multi-layer perceptron was used for color prediction.The experimental results on human dataset shows that this method not only has better performance in objective indicators,but also can ensure a more complete and clear edge point cloud in subjective quality compared with the state-of-the-art methods.