|
Point clouds stand out among other 3D data representations for their effectiveness and adaptability. Thus, point cloud analysis has been a popular research topic in recent years. However, due to complex features, numerous scene categories, occlusion, and noise, reliable semantic segmentation in huge scenes is difficult. Existing point cloud segmentation techniques have two drawbacks: 1) since different nearby points are frequently considered similarly, it is impossible to accurately characterize the relationship between the center point and its vicinity. 2) How to simultaneously address the issues of point cloud semantic segmentation efficiency and data balance in huge scenes. A network paradigm based on the Gated Graph Attention Network (GGAN) is suggested in order to get around these restrictions. To improve local feature extraction, GGAN can highlight not only the significance of various neighbor points but also the significance of various representation Spaces. Recall Loss is employed in the meantime to address the issue of data imbalance brought on by random sampling in order to increase effectiveness. The experimental findings on the Semantic3D dataset revealed that the average intersection ratio of GGAN reached 77.1%, which could ensure the segmentation accuracy of large categories (natural landscape, buildings), as well as small categories (cars), simultaneously. This segmentation performance was superior to that of the existing segmentation methods.
|
