Local fully connected graph-encoded semantic segmentation for point cloud
The fine-grained semantic information of large scenes is playing an important role in city-level 3D real scene construction.In this paper,we propose a local fully connected graph-encoded semantic segmentation network for point cloud.In the network,the local fully connected graph based encoder is used to improve the performance of feature learning in neighborhoods.Then a graph dilated convolution-based residual block is used for feature aggregation in a more efficient manner.To relieve the imbalance in sampling,we use inverse frequency weighting cross entropy as the loss function.In order to verify the effectiveness of the proposed method,the comparative experiments is conducted on the SensatUrban dataset.The result shows that,compared with the existing methods,the proposed method is optimal among various metrics,which proves that our method has effectiveness and practicality in the large-scale scene fine-grained semantic segmentation.
point cloudsemantic segmentationgraph encodingdeep learning3D real scene
NETL
NSTL
万方数据
