点云重建需要处理采集过程中存在的噪声空洞以及不均匀采样等问题。传统方法无法利用大量已有的高质量数据来获取先验信息。本文提出结合多视角投影和隐函数的端到端点云重建方法。首先对点云进行多视角投影得到深度图，基于卷积操作的编码器将深度图转化到到特征空间，并在这个过程中进行去噪和补全。基于多层感知机（MLP，Multi-layer perceptron）实现的隐函数用来表达几何。在训练过程中，对高质量的几何进行点采样，通过点云和点采样的隐函数值来监督学习编码器和隐函数的参数。实验在多个数据集上进行，通过与传统方法以及现有数据驱动的方法进行对比，验证了新方法的有效性，在重建细节、噪声处理和空洞补全均有明显的提高。预训练的模型直接适用于多相机深度采集系统，并能自动权衡去噪和细节保留的程度，在真实数据上重建出光滑的几何表面。

Point cloud reconstruction must deal with the noise, the holes and uneven sampling problems in acquisition. Traditional methods lack the way to learn prior knowledge from large datasets of high-quality geometric models. This paper proposes an end-to-end reconstruction method by combining multi-view projection and implicit function. First the depth map is obtained by projecting the point cloud to multiple views. The convolution-based encoder transforms the depth map into feature space, at the same time denoising and complementing the feature. Implicit function is implemented with Multi-layer Perceptron (MLP), and is used to represent geometry. In the training process, the high-quality mesh is sampled, the encoder and implicit function are supervised by the pairing pointclouds and implicit values of sampling. Experiments are carried on several datasets to show the effectiveness of the new method. By comparing with traditional method and existing data-driven methods, it shows the new method significantly improve the reconstruction details while suppressing scanning noise, and complement the holes in more natural way. The pre-trained model is directly applicable to multi-view depth capturing system, and can automatically balance the degree of denoise and preservation of detail, and reconstruct a visually acceptable geometric surface on real data.

国家自然科学基金项目（面上项目，重点项目，重大项目）