A feature fusion-based attention graph convolutional network for 3D classification and segmentation

Abstract

<abstract><p>Among all usual formats of representing 3D objects, including depth image, mesh and volumetric grid, point cloud is the most commonly used and preferred format, because it preserves the original geometric information in 3D space without any discretization and can provide a comprehensive understanding of the target objects. However, due to their unordered and unstructured nature, conventional deep learning methods such as convolutional neural networks cannot be directly applied to point clouds, which poses a challenge for extracting semantic features from them. This paper proposes a feature fusion algorithm based on attention graph convolution and error feedback, which considers global features, local features and the problem of the features loss during the learning process. Comparison experiments are conducted on the ModelNet40 and ShapeNet datasets to verify the performance of the proposed algorithm, and experimental results show that the proposed method achieves a classification accuracy of 93.1% and a part segmentation mIoU (mean Intersection over Union) of 85.4%. Our algorithm outperforms state-of-the-art algorithms, and effectively improves the accuracy of point cloud classification and segmentation with faster convergence speed.</p></abstract>