Normal Estimation of Surface in PointCloud Data for 3D Parts Segmentation
Takuya Tsujibayashi, Katsufumi Inoue, Michifumi Yoshioka · 2018
3D object parts segmentation is a fundamental task and essential for many higher level tasks, such as object detection, human pose estimation and environment recognition. For 3D object parts segmentation, there are methods using local 3D concave/convex relationships. Those relationships are distinguished by using normals in local adjacent regions of 3D object surface. To obtain these relationships, precise estimation of normals are very important. However, the normal estimation method for a whole surface of an object which are represented by PointCloud is not discussed sufficiently. Therefore, in this research, we propose an automatic normal estimation method. In our method, the number of times that a normal go through surfaces of the object is used to determine the direction of the normal, inside or outside of the object. As results of experiments on the benchmark which consists of 380 3D objects in 19 categories, we obtained the average normal estimation accuracy of 99%.