GAPointNet: Graph attention based point neural network for exploiting local feature of point cloud

Abstract

Exploiting fine-grained semantic features on point cloud data is still challenging because of its irregular and sparse structure in a non-Euclidean space. In order to represent the local feature for each central point that is helpful towards better contextual learning, a max pooling operation is often used to highlight the most important feature in the local region. However, all other geometric local correlations between each central point and corresponding neighbourhood are ignored during the max pooling operation. To this end, the attention mechanism is promising in capturing node representation on graph-based data by attending over all the neighbouring nodes. In this paper, we propose a novel neural network for point cloud analysis, GAPointNet, which is able to learn local geometric representations by embedding graph attention mechanism within stacked Multi-Layer-Perceptron (MLP) layers. Specifically, we highlight different attention weights on the neighbourhood of each center point to efficiently exploit local features. We also combine attention features with local signature features generated by our attention pooling to fully extract local geometric structures and enhance the network robustness. The proposed GAPointNet architecture is tested on various benchmark datasets (i.e. ModelNet40, ShapeNet part, S3DIS, KITTI) and achieves state-of-the-art performance in both the shape classification and segmentation tasks.

Graph Attention-Based Deep Neural Network for 3D Point Cloud Processing

“…Graph attention convolution is basically similar to GAPLayer proposed by Chen et al [17], but there are still some differences. Considering that different neighborhood points have different importance to the center point, the neighborhood attention coefficient is obtained by learning the neighborhood edge features.…”

“…The neighborhood area of each point is changed by using the furthest point sampling to extract features hierarchically. Our network does not use a multi-head mechanism like GAPNet [17], and uses two graph attention convolutions in each feature extraction process, and extracts features in layers. In this way, the stability of the network can be guaranteed, and high-dimensional feature information can also be extracted.…”

“…In Table 1, our method is compared with some other methods such as PointNet++ [5], Kd-Net [11], AdvectiveNet [12]et al Experimental results show that the proposed network achieves relatively good results for the classification task compared with some methods. Compared with GAPNet [17], the classification accuracy is improved by 0.5%. These indicate that our method can fully extract point cloud features, mine potential information, and achieve good recognition results.…”

DNet: Dynamic Neighborhood Feature Learning in Point Cloud

“…Since the points in point cloud are similar to the nodes in graph, some works used graph convolution approaches to process the point cloud [15][16][17][18][19]. The graph convolution approach can be divided into the spectral convolution and the spatial convolution [7].…”

“…At the same time, because the spectral convolution is associated with the Laplacian matrix, its generalization ability is weak. By contrast, the spatial convolution approach can directly perform convolution on the local neighborhoods of point cloud [16][17][18][19], and has high computational efficiency and strong generalization.…”