Point-Cloud Semantic Segmentation Network Considering Normals

In deep learning-based point-cloud semantic classification, PointNet considers the three-dimensional coordinates of the point cloud as a direct input, however, the classification of irregular shape objects is a challenge. In this study, we propose a semantic segmentation network considering the normals of point cloud by adding a normal estimation module on PointNet. We estimate the normals using a principal component analysis method. Compared with the original model, the overall accuracy, mean per-class accuracy, and mean per-class intersection-over-union of the improved model are improved by 2. 3 percentage points, 7. 1 percentage points, and 3. 9 percentage points respectively. Among the 13 semantic classes, the classification accuracy for 10 classes is improved, of which the classification accuracy of sofa and column is improved by 45. 6 percentage points and 42. 2 percentage points, respectively, and the mean per-class intersection-over-union is improved by 19. 8 percentage points and 25. 0 percentage points, respectively. Results show that the semantic segmentation network considering normals can improve the overall performance of the network to a certain extent and can significantly improve the classification effect of sofa and column.