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