Point Cloud Semantic Segmentation for Nursery Based on Multi-Scale Feature Fusion

Fig. 7. Examples of saplings. (a) Labelling of point cloud of the shrub; (b) picture of the shrub; (c) labelling of point cloud of the small tree; (d) picture of the small tree

Download full size

Fig. 8. Overall pictures of the dataset. (a) Original point clouds of the dataset; (b) dataset with label

Download full size

Fig. 9. Partial visualization results for semantic segmentation of the nursery dataset

Download full size

Table 1. The number of points of each category in the dataset
View table
Table 1. The number of points of each category in the dataset
Scene Number of points
crown trunk pot pole ground
1 646630 55541 80776 14769 247169
2 580797 2813 90859 16810 327488
3 805011 33695 86767 46073 286790
4 635954 25483 64417 62105 221914

Table 2. Comparison of the results of the ablation experiments of neighborhood characteristic polymerization uint: %

View table

Table 2. Comparison of the results of the ablation experiments of neighborhood characteristic polymerization uint: %

Method	mIoU	OA	mAcc
Replace with MaxPool	74.46	95.73	88.50
Replace with AttentivePool	86.22	97.33	94.12
Remove multi-head	82.65	96.86	90.13
Remove weights	86.96	97.38	93.96
Remove multi-scale	86.26	97.15	93.04
Proposed method	87.89	97.42	94.27

Table 3. Comparison of experimental results of position encoding blocks
View table
Table 3. Comparison of experimental results of position encoding blocks
Method mIoU OA mAcc
① 76.10 95.56 89.61
② 84.94 97.37 93.13
③ 82.15 96.46 90.97
④ 87.23 97.38 94.00
⑤ 85.16 97.15 93.41
⑥ 87.89 97.42 94.27

Table 4. Comparison of downsampling strategies
View table
Table 4. Comparison of downsampling strategies
Method Time /s Density variance
RS 1×10^‒⁶ 0.31244
FPS 3.94 0.04207
MPS 0.20 0.06616

Table 5. Comparison of experimental results of downsampling strategies
View table
Table 5. Comparison of experimental results of downsampling strategies
Method mIoU OA mAcc
RS 86.23 97.24 93.50
FPS 87.83 97.39 94.13
MPS 87.89 97.42 94.27

Table 6. Experimental results of semantic segmentation on the homemade dataset

View table

Table 6. Experimental results of semantic segmentation on the homemade dataset

Method	mAcc	OA	mIoU	IoU
Method	mAcc	OA	mIoU	crown	trunk	pot	pole	ground
PointNet	53.53	87.09	43.70	92.21	0.88	41.97	8.31	75.13
PointNet++	69.94	92.34	60.05	95.14	30.97	66.07	18.42	89.69
DGCNN	66.55	91.35	56.50	95.10	28.78	58.38	14.39	85.87
RandLA-Net	91.13	96.87	82.80	97.59	75.19	82.90	62.40	95.93
Proposed method	94.27	97.42	87.89	97.61	79.76	85.19	80.52	96.37

Tools

Get Citation

Copy Citation Text

Hui Liu, Siyuan Wang, Jie Xu, Yue Shen, Jinru Kai, Xinpeng Zheng. Point Cloud Semantic Segmentation for Nursery Based on Multi-Scale Feature Fusion[J]. Laser & Optoelectronics Progress, 2025, 62(16): 1615005

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites