Multi-Head Attentional Point Cloud Classification and Segmentation Using Strictly Rotation-Invariant Representations

Point cloud processing plays an increasingly essential role in three-dimensional (3D) computer vision target, scene parsing, environmental perception, etc. Compared with using aligned point cloud data for classification and segmentation, the strictly rotation-invariant representations show enough robustness. Inspired by the great success of deep learning, we propose a novel neural network called Multi-head Attentional Point Cloud Classification and Segmentation Using Strictly Rotation-invariant Representations. Our research focuses on processing the point cloud rotated in any direction effectively and precisely. First of all, the strictly rotation-invariant point cloud representations are obtained through point projection. Then we apply a multi-head attentional convolution layer (MACL) using attention coding to develop the performance of point cloud feature extraction. Finally, our network assigns different responses and recognizes the overall geometry well through a key point descriptor, adding to the global feature. Our method can explore more in-depth information for accuracy enhancement with attention pooling and multi-layer perceptron (MLP) based on an advanced DenseNet. Our network enjoys 90.63% and 87.50% classification accuracy testing on ModelNet10 and ModelNet40, and 75.15% intersection over union metric (mIoU) evaluating on ShapeNet Part dataset, remaining under any rotation. Rotating experimental results indicate that our framework realizes better point cloud classification and segmentation performance than most state-of-the-art methods.