基于多尺度特征融合的由粗到精点云形状补全
Coarse-to-Fine Point Cloud Shape Complementation Based on Multi-Scale Feature Fusion
-
摘要: 为了以由粗到精的方式实现点云形状补全,提出一个端到端的两阶段多尺度特征融合网络,其中的每个阶段都是由一个编码器-解码器构成.第1阶段中,首先利用点集抽取模块提取残缺点云的全局特征,在获取不同分辨率点特征的同时能关注更多的局部邻域特征,然后使用多层感知机作为解码器生成粗糙的点云骨架;第2阶段中,利用点云骨架和残缺点云提取多尺度局部特征,并通过注意力机制与第1阶段中的多尺度全局特征相互融合,使得每个点都包含全局和局部几何信息;最后将第2阶段中的全局特征和多尺度局部特征逐步进行上采样,并通过多层感知机生成精细的完整点云.采用倒角距离作为评价标准,在ShapeNet,MVP和Completion3D数据集上进行点云补全实验的结果表明,误差分别比基准网络降低17.1%,3.9%和13.9%,验证了所提网络的有效性.Abstract: To implement the point cloud shape completion in a coarse-to-fine manner, an end-to-end two-stage multi-scale feature fusion network is proposed, in which each stage consists of an encoder-decoder. In the first stage, the set abstraction module extracts the global features of the incomplete point cloud, which can focus on more local neighborhood features while acquiring point features of different resolutions. A decoder built from multilayer perceptrons generates a coarse skeleton. In the second stage, the coarse skeleton and incomplete point cloud are used to learn multi-scale local features. The multi-scale local features are fused with the multi-scale global features of the first stage via an attention mechanism so that each point contains global and local geometric information. Finally, the global features and multi-scale local features are progressively upsampled, and a fine-grained complete point cloud is generated via multilayer perceptrons. The point cloud completion experiments on ShapeNet, MVP, and Completion3D datasets show that the chamfer distance is reduced by 17.1%, 3.9%, and 13.9% compared with the baseline, respectively, demonstrating the effectiveness of the proposed method.