摘要：

利用隐函数对非结构化点云进行物体重建，具有不受拓扑结构和分辨率限制的特点，但是存在重建物体结构不准确、缺乏局部细节等问题。为此，提出一种样条编码和长短期记忆(LSTM) 网络相结合的改进点云物体重建算法。针对重建物体缺乏局部细节的问题，提出一种增强细节表示能力的样条编码模块，采用均匀二次B样条提取点云的局部位置特征信息。为了提高隐函数的重建准确率，设计LSTM点云预测模块，挖掘点云潜在的空间结构信息，利用包含空间相关性的动态权重自适应地预测靠近物体表面的点云，进一步提高隐式重建的准确率。为提高点云预测的合理性，引入点云变形损失和结构排斥损失，用于优化预测点云的整体结构。在2个公开的三维模型数据集ShapeNet和ABC上进行实验验证，结果表明，该算法的倒角距离分别为0.009%和0.427%，与DeepSDF、CurrSDF和MetaSDF算法相比，所提算法能够重建出更丰富的细节特征，获得更准确的重建结果及良好的视觉效果；在真实世界数据集Pix3D上的实验结果表明，该算法相较对比算法具有更优的泛化性能。

关键词: 点云重建, 隐式重建, 深度学习, 样条编码, 长短期记忆网络

Abstract:

Using an implicit function to reconstruct objects from unstructured point clouds is not limited by topology and resolution, but there are some problems, such as inaccurate structure of reconstructed objects and lack of local details. Therefore, an improved point cloud object reconstruction algorithm based on spline coding and Long Short-Term Memory (LSTM) is proposed. A spline encoding module is proposed to enhance the ability of detail representation to address the problem of lacking local details in reconstructed objects. Uniform quadratic B-splines are used to extract local position feature information of point clouds. To improve the reconstruction accuracy of the implicit function, the LSTM point cloud prediction module is designed to mine the potential spatial structure information of the point cloud. The dynamic weight containing spatial correlation is used to adaptively predict the point cloud near the object surface, further improving the accuracy of implicit reconstruction. To improve the rationality of point cloud prediction, point cloud deformation loss and structural rejection loss are introduced to optimize the overall structure of the predicted point cloud. Experimental validation was conducted on two publicly available 3D model datasets, ShapeNet and ABC, and the results showed that the chamfer distances of the algorithm were 0.009% and 0.427%, respectively. Compared with the DeepSDF, CurrSDF, and MetaSDF algorithms, the proposed algorithm can reconstruct richer detailed features, obtain more accurate reconstruction results, and achieve good visual effects. Experimental results on the real-world dataset Pix3D show that this algorithm has better generalization performance than the comparison algorithms.

Key words: point cloud reconstruction, implicit reconstruction, deep learning, spline encoding, Long Short-Term Memory(LSTM) network