融合门控自校准机制和图卷积网络的点云分析

Point Cloud Analysis Combining Gated Self-Calibration Mechanism and Graphical Convolutional Network

与密集网格表示的图像不同,点云自身具有不规则和无序性的特点,因而如何准确地推理出点云数据中的形状特征是一项具有挑战性的工作。为解决当前研究存在的不足,提出了点集内-外形状卷积(IE-Conv)。该卷积利用高效的双边结构将点集内的局部形状与点集外的全局形状分开处理,在点集内部,基于门控的方式有选择地学习丰富的点间关系,同时利用自校准功能优化逐点与局部特征;在点集外部,通过图卷积构建全局图形,并聚焦于点集之间的远程依赖关系;最后将双边输出有机融合起来。将IE-Conv分层嵌入形状推理卷积网络中(SR-Net),并在标准ModelNet40和ShapeNet数据集上进行分类与分割实验。实验结果表明,分类任务精度达到93.9%,分割任务平均交并比达到86.4%,验证了SR-Net在点云分析中的良好性能。

Point clouds, unlike images represented by dense grids, are characterized by irregularity and disorder, making it difficult to precisely reason out the shape features in point cloud data. The internal-external shape son volution for point sets(IE-Conv) is proposed to address the limitations of current research. The local shape inside the point set is treated separately from the global shape outside the point set using an efficient bilateral design. Rich inter-point relationships are selectively studied in a gate-based manner within the point set, while point-by-point and local features are optimized by self-calibration functions;outside the point set, global shapes are constructed using graph convolution and focus on longrange dependencies between point sets. Finally, the organic fusion of the bilateral outputs is performed. This paper performs classification and segmentation experiments on the standard ModelNet40 and ShapeNet datasets by hierarchically embedding IE-Conv into the shape-reasoning convolutional network(SR-Net). The experimental results show that the classification task achieves an accuracy of 93. 9% and the segmentation task achieves the mean intersection over the union of 86. 4%, which verifies the good performance of SR-Net in point cloud analysis.