Abstract: Point cloud completion aims to reconstruct a high-quality complete point cloud from incomplete point cloud data. However, most existing point cloud completion networks have limitations in capturing local features and reconstructing details, resulting in poor performance in terms of local details and completion accuracy. To address these issues, a multi-stage point cloud completion Network based on Adaptive Neighborhood Feature Fusion (ANFF-Net) was proposed. Firstly, the neighborhood selection of key points was adjusted by the feature extractor adaptively to adapt to different shapes of point clouds, effectively capturing the spatial relationships between points with different semantics, thereby reducing the loss of local details. Then, a local perception Transformer was used by the feature expander to further expand the local feature information of neighboring points, improving the network's ability to recover details. Finally, a cross-attention mechanism was applied by the point cloud generator to selectively propagate the local feature information of the incomplete point cloud and a folding module was used to gradually refine the local regions, significantly enhancing the detail retention of the completed point cloud and generating more consistent geometric details.  Experimental results show that ANFF-Net improves the completion accuracy by 9.67 percentage points compared to ProxyFormer on the ShapeNet55 dataset and achieves good completion performance on the PCN and KITTI datasets. The visualization results indicate that the point clouds generated by ANFF-Net have finer granularity and are closer in shape to the ground truth.

Key words: point cloud completion, local feature, adaptive neighborhood, local perception, cross-attention

摘要： 点云补全指利用不完整的点云数据重建高质量的完整点云。然而，现有的大多数点云补全网络在捕捉局部特征和细节重建方面存在不足，导致生成的点云在局部细节和补全精度上表现不佳。为解决上述问题，提出了一种基于自适应邻域特征融合的多阶段点云补全网络(ANFF-Net)。首先，特征提取器通过自适应调整关键点的邻域选择，以适应不同形状的点云，有效捕捉不同语义点之间的空间关系，减少局部细节信息的丢失；其次，特征拓展器利用局部感知Transformer进一步扩展邻近点的局部特征信息，提升网络的细节恢复能力；最后，点云生成器采用交叉注意力机制，选择性传递不完整点云的局部特征信息，并使用折叠模块逐步细化点云的局部区域，显著增强补全后点云的细节保留，生成更一致的几何细节。实验结果表明，ANFF-Net在ShapeNet55数据集上的补全精度相较于ProxyFormer提升了9.67%，并在PCN和KITTI数据集上也取得了较好的补全效果；可视化结果显示，ANFF-Net生成的点云具有更高的细粒度，形状更接近真实值。

关键词: 点云补全, 局部特征, 自适应邻域, 局部感知, 交叉注意力