An End-to-End Fine-Grained Classification Network for 3D Point Clouds

In terms of the problem of insufficient interpretation for sub-categories in meta-class recognition by 3D point clouds classification algorithms based on deep learning,a 3D model classification framework oriented to point clouds is proposed to build an end-to-end fine-grained point clouds network FGP-Net, which contains inter-layer semantic correlation and intra-layer context-awareness. That is, dense connections blocks are constructed by interconnected convolution operators, and local regions are constructed by the ball query in the layers to generate local regions for completing the feature mapping from local to global, then it passes through the offset-attention mechanism to pay attention to the contextual differences, so as to better capture fine-grained attributes. The multi-layer feature aggregation strategy is applied to explore the correlation between the features of each layer, and the relevant semantic information is learned through back-propagation to improve the classification performance of the model. On three sub-datasets of FG3D, Airplane, Chair and Car, the overall accuracy reaches 95.77%, 80.88% and 77.94% respectively. Compared with the advanced 3D point clouds classification models PointNet++, PointCNN, Point2Sequence and DGCNN etc, FGP-Net has certain advantages in classification performance.