Fig. 1. Fine-grained annotations of 24 objects in the PartNet dataset
[26] Fig. 2. Models of chairs and laptops in the ShapeNet Part dataset
[28] Fig. 3. Point cloud scene and semantic segmentation diagram in S3DIS dataset
[29] Fig. 4. Annotated indoor scene maps in ScanNet dataset
[30] Fig. 5. Examples of Semantic3D dataset
[31]. (a) Point cloud scene; (b) diagram of intensity; (c) diagram of semantic segmentation
Fig. 6. Semantic segmentation results of outdoor scene in vKITTI dataset
[32] Fig. 7. Visual representation of point cloud semantic segmentation methods
Fig. 8. Framework of multi-view convolutional neural network (MVCNN)
[37] Fig. 9. Overall flow of semantic marks of SnapNet-R network
[39] Fig. 10. Overall architecture of semantic segmentation network of SEGCloud
[49] Fig. 11. Framework of PointNet for point cloud classification and segmentation
[16] Fig. 12. Overall framework of RSNet for point cloud semantic segmentation
[53] Fig. 13. Point cloud semantic segmentation network of SO-Net formed by SOM
[54] Fig. 14. Applications of hierarchical convolution in regular gird and point clouds, and PointCNN framework used for semantic segmentation
[18]. (a) Application of hierarchical convolution; (b) PointCNN framework
Fig. 15. Architecture of PointNet++ for point cloud classification and segmentation
[17] Fig. 16. Overall architectures of PointSIFT module and point segementation of PointSIFT
[58]. (a) Structure; (b) whole architecture
Fig. 17. Architecture of A-CNN for point cloud classification and segmentation
[61] Fig. 18. Point cloud semantic segmentation network of 3DMAX-Net
[60] (MS-FLB: multi-scale feature learning block; LGAB: local and global feature aggregation block)
Fig. 19. Framework of 3P-RNN for point cloud semantic segmentation
[62] Fig. 20. Network structural diagram of LDGCNN for point cloud classification and segmentation
[65] Fig. 21. Network structural diagram of RGCNN for point cloud classification and segmentation
[67] Fig. 22. Network framework of GAPNet for point cloud semantic segmentation
[68] Fig. 23. Forward time of different network models
Dataset | Number ofcategories | Number oftraining sets | Number ofverification sets | Number oftesting sets |
---|
PartNet[26] | 24 | - | - | - | UWA Dataset[27] | 55 | - | - | - | ShapeNet Part[28] | 16 | 12137 | 1870 | 2874 | S3DIS[29] | 13 | 224 | - | 48 | ScanNet[30] | 21 | 1201 | - | 312 | Semantic3D[31] | 8 | 15 | - | 15 | KITTI(Zhang)[34] | 10 | 140 | - | 112 | KITTI(Ros)[35] | 11 | 170 | - | 46 | vKITTI[32] | 13 | - | - | - |
|
Table 1. Common datasets of point cloud segmentation
Model | Network parameter quantity /MB |
---|
Subvolume | 16.6 | MVCNN | 60 | PointNet | 3.48 | PointNet++ | 1.48 | DGCNN | 1.84 | LDGCNN | 1.08 | PointCNN | 0.6 |
|
Table 2. Network parameter quantity of different semantic segmentation models
Model | S3DIS | ScanNet | ShapeNet Part | Semantic 3D | vKITTI | | | | |
---|
mIoU | OA /% | mIoU | OA /% | | | | | mIoU | mIoU | OA /% | mIoU | OA /% |
---|
Yi[69] | - | - | - | - | 81.4 | - | - | - | - | KD-Net[47] | - | - | - | - | 82.3 | - | - | - | - | SEGCloud[49] | 48.92 | - | - | 73.0 | 79.4 | 61.3 | 88.1 | - | - | PointNet[16] | 47.71 | 78.62 | 14.69 | 73.9 | 83.7 | - | - | 34.4 | 79.7 | PN++(SSG)[17] | - | - | - | 83.3 | - | - | - | - | - | PN++(MSG+DP)[17] | - | - | 34.26 | 84.5 | 85.1 | - | - | - | - | PN++(MRG+DP)[17] | - | - | - | 83.4 | - | - | - | - | - | O-CNN+CRF[70] | - | - | - | - | 85.9 | - | - | - | - | SSCNN[71] | - | - | - | - | 84.7 | - | - | - | - | MS+CU[72] | 47.8 | 79.2 | - | - | - | - | - | - | - | G+RCU[72] | 49.7 | 81.1 | - | - | - | - | - | 36.2 | 80.6 | DGCNN[64] | 56.1 | 84.1 | - | - | 85.1 | - | - | - | - | RGCNN[67] | - | - | - | - | 84.3 | - | - | - | - | RSNet[53] | 53.83 | 61.81 | 39.35 | - | 84.9 | - | - | - | - | SO-Net[54] | - | - | - | - | 84.6 | - | - | - | - | TMLC-MSR[73] | - | - | - | - | - | 54.2 | 86.2 | - | - | DeePr3SS[74] | - | - | - | - | - | 58.5 | 88.9 | - | - | SnapNet[38] | - | - | - | - | - | 59.1 | 88.6 | - | - | SGPN[75] | 50.37 | 80.78 | - | - | 85.8 | - | - | - | - | SpiderCNN[59] | - | - | - | - | 85.3 | - | - | - | - | 3DMAX-Net[60] | 47.5 | 79.5 | - | - | - | - | - | - | - | SPGraph[76] | 62.1 | 85.5 | - | - | - | 73.2 | 94.0 | - | - | 3P-RNN[62] | 56.3 | 86.9 | - | - | - | - | - | 41.6 | 87.8 | PointCNN[18] | 62.74 | 88.1 | - | 85.1 | - | - | - | - | - | PointSIFT[58] | 70.23 | 88.72 | - | 86.2 | - | - | - | - | - | ASIS[77] | 59.3 | 86.2 | - | - | - | - | - | - | - | A-CNN[61] | - | 87.3 | - | - | - | - | - | - | - | LDGCNN[65] | - | - | - | - | 85.1 | - | - | - | - | GAPNet[68] | - | - | - | - | 84.7 | - | - | - | - |
|
Table 3. Segmentation results of different models on typical point cloud datasets