摘要:
受到三维激光扫描仪分辨率限制、环境遮挡等因素的影响, 扫描获取的树木点云通常存在缺失, 尤其是单视角扫描的点云缺失更为严重. 根据树木复杂的几何结构, 提出一种树木点云补全网络. 利用交叉注意力和自注意力机制充分学习输入树木缺失点云的潜在特征, 并通过解码该特征从稀疏到精细地预测完整树木点云; 针对树木补全数据集难以获取的问题, 采用模拟单视角扫描的方式构建有真值且包括不同树木类型的树木点云补全数据集. 在所构建的数据集上的实验结果表明, 与点云补全网络AdaPoinTr相比, 所提网络的平均倒角距离降低0.62, 平均F分数增加0.04; 该网络可以有效地补全不同种类的缺失树木点云.
Abstract:
Due to the limitations of three-dimensional laser scanners, such as resolution and environmental occlusion, the scanned point clouds of trees are usually incomplete, especially for single view scans. In light of the complex geometric structures of trees, a tree point cloud completion network is proposed in this paper, which fully learns the potential features of missing point clouds through cross-attention and self-attention mechanisms, and predicts complete tree point clouds from sparse to refined by decoding those features; To address the problem of difficult acquisition of tree completion datasets, a tree point cloud completion dataset with ground truth from different tree type using simulated scans of the single view is constructed in this paper. Experiments on this dataset show that compared to the point cloud completion network AdaPoinTr, the proposed network in this paper reduces the average Chamfer distance by 0.62 and increases the average F-score by 0.04; Thus, the proposed network can effectively complete missing point clouds of trees of different types.
下载:
