In the 3D point cloud semantic segmentation algorithm based on deep learning, to enhance the fine-grained ability to extract local features and learn the long-range dependencies between different local neighborhoods, a neural network based on attention mechanism and global feature optimization was proposed. First, a Single-Channel Attention (SCA) module and a Point Attention (PA) module were designed in the form of additive attention. The former strengthened the resolution of local features by adaptively adjusting the features of each point in a single channel, and the latter adjusted the importance of the single-point feature vector to suppress useless features and reduce feature redundancy. Second, a Global Feature Aggregation (GFA) module was added to aggregate local neighborhood features to capture global context information, thereby improving semantic segmentation accuracy. The experimental results show that the proposed network improves the mean Intersection?over?Union (mIoU) by 1.8 percentage points compared with RandLA-Net (Random sampling and an effective Local feature Aggregator Network) on the point cloud dataset S3DIS, and has good segmentation performance and good adaptability.
Graph Neural Network (GNN) is vulnerable to adversarial attacks, leading to performance degradation, which affects downstream tasks such as node classification, link prediction and community detection. Therefore, the defense methods of GNN have important research value. Aiming at the problem that GNN has poor robustness when being adversarially attacked, taking Graph Convolutional Network (GCN) as the model, an improved Singular Value Decomposition (SVD) based poisoning attack defense method was proposed, named ISVDatt. In the poisoning attack scenario, the attacked graph was able to be purified by the proposed method. When the GCN was attacked by poisoning, the connected edges with large different features were first screened and deleted to keep the graph features smooth. Then, SVD and low-rank approximation operations were performed to keep the low rank of the attacked graph and clean it up. Finally, the purified graph was used for training GCN model to achieve effective defense against poisoning attack. Experiments against Metattack and DICE were conducted on the open source datasets such as Citeseer, Cora and Pubmed, and compared with the defense methods based on SVD, Pro_GNN and Robust Graph Convolutional Network (RGCN), respectively. The results show that ISVDatt has relatively better defense effect, although the classification accuracy is lower than that of Pro_GNN, but it has low complexity and negligible time overhead. Experimental results verify that ISVDatt can resist poisoning attack effectively with the consideration of both the complexity and versatility of the algorithm, and has a high practical value.
Group activity recognition is a challenging task in complex scenes, which involves the interaction and the relative spatial position relationship of a group of people in the scene. The current group activity recognition methods either lack the fine design or do not take full advantage of interactive features among individuals. Therefore, a network framework based on partitioned attention mechanism and interactive position relationship was proposed, which further considered individual limbs semantic features and explored the relationship between interaction feature similarity and behavior consistency among individuals. Firstly, the original video sequences and optical flow image sequences were used as the input of the network, and a partitioned attention feature module was introduced to refine the limb motion features of individuals. Secondly, the spatial position and interactive distance were taken as individual interaction features. Finally, the individual motion features and spatial position relation features were fused as the features of the group scene undirected graph nodes, and Graph Convolutional Network (GCN) was adopted to further capture the activity interaction in the global scene, thereby recognizing the group activity. Experimental results show that this framework achieves 92.8% and 97.7% recognition accuracy on two group activity recognition datasets (CAD (Collective Activity Dataset) and CAE (Collective Activity Extended Dataset)). Compared with Actor Relationship Graph (ARG) and Confidence Energy Recurrent Network (CERN) on CAD dataset, this framework has the recognition accuracy improved by 1.8 percentage points and 5.6 percentage points respectively. At the same time, the results of ablation experiment show that the proposed algorithm achieves better recognition performance.