1, 3D point cloud segmentation needs to know the global geometric structure and fine-grained details of each point. According to the granularity of segmentation, 3D point cloud segmentation methods can be divided into three categories: semantic segmentation (scene level), instance segmentation (object level) and component segmentation (component level).
2. Semantic segmentation of 3D point clouds has applications in many fields, such as autonomous driving, robots and so on. At present, it has become the key to scene understanding.
3. 3D point cloud image annotation data is the basic training data of unmanned driving technology. Three-dimensional point cloud image labeling is to label the target objects, including vehicles, pedestrians, advertising signs, trees and so on, through three-dimensional frames in the three-dimensional images collected by lidar.
4. 3D point cloud continuous frame labeling is a widely used data processing type in autonomous driving scene, which requires high 3D spatial perception and multi-frame collaborative processing ability. Three-dimensional data can usually be expressed in different formats, including depth image, point cloud, grid and volume grid. As a common format, point cloud means that the original geometric information remains in three-dimensional space without any discretization.