Multi-scale supervoxel feature aggregation network for three-dimensional point cloud semantic segmentation

doi:10.11947/j.AGCS.2026.20260064

Abstract

Abstract:

Existing point cloud Transformer module equate spatial proximity with semantic similarity. While valid on continuous surfaces, this assumption often fails at physical boundaries where spatially adjacent points belong to distinct objects, leading to the challenge of boundary feature blurring. This deficiency stems from the inherent limitations of standard KNN grouping. To better extract the semantic information of point clouds, we propose multi-scale supervoxel feature aggregation network (MS-SFA-Net), which injects geometric constraints directly into the encoder. This network replaces simple aggregation with a dual-layer masked attention mechanism. Specifically, it strictly confines intra-supervoxel attention within supervoxels generated by VCCS, and subsequently utilizes inter-supervoxel attention to capture global relationships among supervoxels. On ScanNet v2, the module exhibits significant generalization capability: achieving gains of 1.2% over the Point Transformer V3 baseline and 10.8% over PointNet++, validating its robustness across different backbone networks. Moreover, on the test set of the outdoor Semantic KITTI dataset, our network achieves 1.1% performance improvement over the baseline model, demonstrating its generalization efficacy across diverse application scenarios.

Key words: semantic segmentation, point cloud, supervoxel, deep learning, indoor 3D

CLC Number:

P237

Xijiang Chen, Minkun Zeng, Wei Xuan, Jingui Zou, Xianghong Hua. Multi-scale supervoxel feature aggregation network for three-dimensional point cloud semantic segmentation[J]. Acta Geodaetica et Cartographica Sinica, 2026, 55(8): 1414-1424.

Add to citation manager EndNote|Reference Manager|ProCite|BibTeX|RefWorks

URL: http://xb.chinasmp.com/EN/10.11947/j.AGCS.2026.20260064

http://xb.chinasmp.com/EN/Y2026/V55/I8/1414

Figures/Tables 13

Fig. 1

Fig. 2

Fig. 3

Tab. 1

Tab. 2

Tab. 3

Fig. 4

Tab. 4

Tab. 5

Tab. 6

Fig. 5

Tab. 7

Tab. 8

References 42

[1]	杨必胜, 董震. 点云智能研究进展与趋势[J]. 测绘学报, 2019, 48(12): 1575-1585. DOI: . doi: 10.11947/j.AGCS.2019.20190465
	Yang Bisheng, Dong Zhen. Progress and perspective of point cloud intelligence[J]. Acta Geodaetica et Cartographica Sinica, 2019, 48(12): 1575-1585. DOI: . doi: 10.11947/j.AGCS.2019.20190465
[2]	Li Jie, Liu Zhao, Li Li, et al. Multi-view convolutional vision transformer for 3D object recognition[J]. Journal of Visual Communication and Image Representation, 2023, 95: 103906.
[3]	Feng Yifan, Zhang Zizhao, Zhao Xibin, et al. GVCNN: group-view convolutional neural networks for 3D shape recognition[C]//Proceedings of 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Salt Lake City: IEEE, 2018: 264-272.
[4]	徐涛, 杨元维, 高贤君, 等. 融合图卷积与多尺度特征的接触网点云语义分割[J]. 测绘学报, 2024, 53(8): 1624-1633. DOI: . doi: 10.11947/j.AGCS.2024.20230198
	Xu Tao, Yang Yuanwei, Gao Xianjun, et al. Integrated graph convolution and multi-scale features for the overhead catenary system point cloud semantic segmentation[J]. Acta Geodaetica et Cartographica Sinica, 2024, 53(8): 1624-1633. DOI: . doi: 10.11947/j.AGCS.2024.20230198
[5]	Wu Z, Song S, Khosla A, et al. 3D ShapeNets: a deep representation for volumetric shapes[C]//Proceedings of 2015 IEEE Conference on Computer Vision and Pattern Recognition. Boston: IEEE, 2015: 1912-1920.
[6]	Meng H Y, Gao L, Lai Y, et al. VV-Net: voxel VAE net with group convolutions for point cloud segmentation[C]//Proceedings of 2019 IEEE/CVF International Conference on Computer Vision. Seoul: IEEE, 2019: 8499-8507.
[7]	Qi C R, Su H, Mo K, et al. PointNet: deep learning on point sets for 3D classification and segmentation[C]//Proceedings of 2017 IEEE Conference on Computer Vision and Pattern Recognition. Honolulu: IEEE, 2017: 77-85.
[8]	Qi C R, Yi L, Su H, et al. PointNet++: deep hierarchical feature learning on point sets in a metric space[C]//Proceedings of 2017 Advances in Neural Information Processing Systems. Red Hook: Curran Associates Inc., 2017: 5105-5114.
[9]	Jiang Mingyang, Wu Yiran, Zhao Tianqi, et al. PointSIFT: a SIFT-like network module for 3D point cloud semantic segmentation[PP/OL]. V2. arXiv (2018-11-24) [2026-01-02]. https://doi.org/10.48550/arXiv.1807.00652.
[10]	Zhao H, Jiang L, Fu C W, et al. PointWeb: enhancing local neighborhood features for point cloud processing[C]//Proceedings of 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Long Beach: IEEE, 2019: 5560-5568.
[11]	张蕊, 包丰璞, 黄官龙, 等. 联合多尺度扩展融合与动态残差的街景点云语义分割[J/OL]. 计算机工程与应用, 2025-09-19. https://link.cnki.net/urlid/11.2127.tp.20250918.1853.004.
	Zhang Rui, Bao Fengpu, Huang Guanlong, et al. Semantic segmentation of street scene point clouds by combining multi-scale extended fusionand dynamic residual[J/OL]. China Industrial Economics, 2025-09-19. https://link.cnki.net/urlid/11.2127.tp.20250918.1853.004.
[12]	Vaswani A, Shazeer N, Parmar N, et al. Attention is all you need[C]//Proceedings of 2017 Advances in Neural Information Processing Systems: Red Hook: Curran Associates Inc., 2017: 5998-6008.
[13]	库闵凡, 张良, 邓继伟, 等. LAE-Transformer：融合局部感知增强的机载LiDAR点云分割网络[J]. 遥感学报, 2025, 29(11): 3363-3377.
	Ku Minfan, Zhang Liang, Deng Jiwei, et al. LAE-Transformer: an airborne LiDAR point cloud segmentation network incorporating local-aware enhancement[J]. National Remote Sensing Bulletin, 2025, 29(11): 3363-3377.
[14]	Guo Menghao, Cai Junxiong, Liu Zhengning, et al. PCT: point cloud transformer[J]. Computational Visual Media, 2021, 7(2): 187-199.
[15]	Zhao Hengshuang, Jiang Li, Jia Jiaya, et al. Point transformer[C]//Proceedings of 2021 IEEE/CVF International Conference on Computer Vision. Montreal: IEEE, 2021: 16239-16248.
[16]	Liu Zheng, Zhou Weijie, Guo Chuchen, et al. PyramidPCD: a novel pyramid network for point cloud denoising[J]. Pattern Recognition, 2025, 161: 111228.
[17]	Lai Xin, Liu Jianhui, Jiang Li, et al. Stratified transformer for 3D point cloud segmentation[C]//Proceedings of 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition. New Orleans: IEEE, 2022: 8490-8499.
[18]	Zhou Wei, Zhao Yiheng, Xiao Yi, et al. TNPC: Transformer-based network for point cloud classification[J]. Expert Systems with Applications, 2024, 239: 122438.
[19]	Jiang Chenru, Huang Kaizhu, Wu Junwei, et al. PointGS: bridging and fusing geometric and semantic space for 3D point cloud analysis[J]. Information Fusion, 2023, 91: 316-326.
[20]	Wu Junwei, Sun Mingjie, Jiang Chenru, et al. Context-based local-global fusion network for 3D point cloud classification and segmentation[J]. Expert Systems with Applications, 2024, 251: 124023.
[21]	Cheng Haozhe, Zhu Jihua, Lu Jian, et al. EDGCNet: joint dynamic hyperbolic graph convolution and dual squeeze-and-attention for 3D point cloud segmentation[J]. Expert Systems with Applications, 2024, 237: 121551.
[22]	Yu Xumin, Tang Lulu, Rao Yongming, et al. Point-BERT: pre-training 3D point cloud transformers with masked point modeling[C]//Proceedings of 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition. New Orleans: IEEE, 2022: 19291-19300.
[23]	Pang Y, Wang W, Tay F E H, et al. Masked autoencoders forPoint cloud self-supervised learning[C]//Proceedings of 2022 European Conference on Computer Vision. Cham: Springer, 2022: 604-621.
[24]	Min Chen, Xiao Liang, Zhao Dawei, et al. Occupancy-MAE: self-supervised pre-training large-scale LiDAR point clouds with masked occupancy autoencoders[J]. IEEE Transactions on Intelligent Vehicles, 2024, 9(7): 5150-5162.
[25]	Hu Qingyong, Yang Bo, Xie Linhai, et al. RandLA-Net: efficient semantic segmentation of large-scale point clouds[C]//Proceedings of 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Seattle: IEEE, 2020: 11105-11114.
[26]	Landrieu L, Simonovsky M. Large-scale point cloud semantic segmentation with superpoint graphs[C]//Proceedings of 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Salt Lake City: IEEE, 2018: 4558-4567.
[27]	Papon J, Abramov A, Schoeler M, et al. Voxel cloud connectivity segmentation-supervoxels for point clouds[C]//Proceedings of 2013 IEEE Conference on Computer Vision and Pattern Recognition. Portland: IEEE, 2013: 2027-2034.
[28]	Dai A, Chang A X, Savva M, et al. ScanNet: richly-annotated 3D reconstructions of indoor scenes[C]//Proceedings of 2017 IEEE Conference on Computer Vision and Pattern Recognition. Honolulu: IEEE, 2017: 2432-2443.
[29]	Behley J, Garbade M, Milioto A, et al. SemanticKITTI: a dataset for semantic scene understanding of LiDAR sequences[C]//Proceedings of 2019 IEEE/CVF International Conference on Computer Vision. Seoul: IEEE, 2019: 9296-9306.
[30]	Wu Xiaoyang, Jiang Li, Wang Pengshuai, et al. Point transformer V3: simpler, faster, stronger[C]//Proceedings of 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Seattle: IEEE, 2024: 4840-4851.
[31]	Dai A, Nießner M. 3DMV: joint 3D-multi-view prediction for 3D semantic scene segmentation[C]//Proceedings of Computer Vision-ECCV2018. Cham: Springer, 2018: 458-474.
[32]	Song Hamin, Cho J, Ha Jinsu, et al. Panoptic-FusionNet: camera-LiDAR fusion-based point cloud panoptic segmentation for autonomous driving[J]. Expert Systems with Applications, 2024, 251: 123950.
[33]	Li Yangyan, Bu Rui, Sun Mingchao, et al. PointCNN: convolution on X-transformed points[PP/OL]. V5. arXiv (2018-11-05) [2026-01-02]. https://arxiv.org/pdf/1801.07791.
[34]	Wu Wenxuan, Qi Zhongang, Li Fuxin. PointConv: deep convolutional networks on 3D point clouds[C]//Proceedings of 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Long Beach: IEEE, 2019: 9613-9622.
[35]	Zhao Lin, Tao Wenbing. JSNet: joint instance and semantic segmentation of 3D point clouds[C]//Proceedings of the 34th AAAI Conference on Artificial Intelligence/the 32nd Innovative Applications of Artificial Intelligence Conference/the 10th AAAI Symposium on Educational Advances in Artificial Intelligence New York: [s.n.], 2020: 12951-12958.
[36]	Yan Xu, Zheng Chaoda, Li Zhen, et al. PointASNL: robust point clouds processing using nonlocal neural networks with adaptive sampling[C]//Proceedings of 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Seattle: IEEE, 2020: 5588-5597.
[37]	Lei H, Akhtar N, Mian A. SegGCN: efficient 3D point cloud segmentation with fuzzy spherical kernel[C]//Proceedings of 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Seattle: IEEE, 2020: 11608-11617.
[38]	Thomas H, Qi C R, Deschaud J E, et al. KPConv: flexible and deformable convolution for point clouds[C]//Proceedings of 2019 IEEE/CVF International Conference on Computer Vision. Seoul: IEEE, 2019: 6410-6419.
[39]	Graham B, Engelcke M, van der Maaten L. 3D semantic segmentation with submanifold sparse convolutional networks[C]//Proceedings of 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Salt Lake City: IEEE, 2018: 9224-9232.
[40]	Choy C, Gwak J, Savarese S. 4D spatio-temporal ConvNets: Minkowski convolutional neural networks[C]//Proceedings of 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Long Beach: IEEE, 2019: 3070-3079.
[41]	Tatarchenko M, Park J, Koltun V, et al. Tangent convolutions for dense prediction in 3D[C]//Proceedings of 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Salt Lake City: IEEE, 2018: 3887-3896.
[42]	Tang Haotian, Liu Zhijian, Zhao Shengyu, et al. Searching efficient 3D architectures with sparse point-voxel convolution[C]//Proceedings of Computer Vision-ECCV2020. Cham: Springer, 2020: 685-702.

方法	方法类别	验证	测试
PointNet++^[8]	点	53.5	55.7
3DMV^[31]	点	－	48.4
PanopticFusion^[32]	点	－	52.9
PointCNN^[33]	点	－	45.8
PointConv^[34]	点	61.0	66.6
JointPointBased^[35]	点	69.2	63.4
PointASNL^[36]	点	63.5	66.6
SegGCN^[37]	点	－	58.9
RandLA-Net^[25]	点	－	64.5
KPConv^[38]	点	－	68.4
SparseConvNet^[39]	体素	69.3	72.5
MinkowskiNet^[40]	体素	72.2	73.6
PTv3^[30]	点	76.8	73.6
PointNet+++MS-SFA	点+超体素	64.3	67.5
PTv3+MS-SFA	点+超体素	78.0	74.1

方法	墙	地板	柜子	床	椅子	沙发	桌子	门	窗	书架	画板	内阁	书桌	窗帘	冰箱	浴帘	马桶	水池	浴缸	其他
PTv3^[30]	87.4	95.8	71.7	83.2	92.4	82.3	77.5	74.7	72.8	82.6	41.8	69.3	70.8	77.3	68.3	71.8	93.3	70.4	89.0	63.2
PTv3+MS-SFA	87.8	95.9	71.5	84.2	92.4	84.9	80.2	75.0	74.6	81.6	41.4	70.9	74.2	78.9	71.0	75.3	94.6	69.8	89.3	66.7

方法	浴缸	床	书架	柜子	椅子	内阁	窗帘	书桌	门	地板	其他	画	冰箱	浴帘	水池	沙发	桌子	马桶	墙	窗
PointNet++^[8]	73.5	66.1	68.6	49.1	74.4	39.2	53.9	45.1	37.5	94.6	37.6	20.5	40.3	35.6	55.3	64.3	49.7	82.4	75.6	51.5
3DMV^[31]	48.4	53.8	64.3	42.4	60.6	31.0	57.4	43.3	37.8	79.6	30.1	21.4	53.7	20.8	47.2	50.7	41.3	69.3	60.2	53.9
PanopticFusion^[32]	49.1	68.8	60.4	38.6	63.2	22.5	70.5	43.4	29.3	81.5	34.8	24.1	49.9	66.9	50.7	64.9	44.2	79.6	60.2	56.1
PointCNN^[33]	57.7	61.1	35.6	32.1	71.5	29.9	37.6	32.8	31.9	94.4	28.5	16.4	21.6	22.9	48.4	54.5	45.6	75.5	70.9	47.5
PointConv^[34]	78.1	75.9	69.9	64.4	82.2	47.5	77.9	56.4	50.4	95.3	42.8	20.3	58.6	75.4	66.1	75.3	58.8	90.2	81.3	64.2
JointPointBased^[35]	61.4	77.8	66.7	63.3	82.5	42.0	80.4	46.7	56.1	95.1	49.4	29.1	56.6	45.8	57.9	76.4	55.9	83.8	81.4	59.8
PointASNL^[36]	70.3	78.1	75.1	65.5	83.0	47.1	76.9	47.4	53.7	95.1	47.5	27.9	63.5	69.8	67.5	75.1	55.3	81.6	80.6	70.3
SegGCN^[37]	83.3	73.1	53.9	51.4	78.9	44.8	46.7	57.3	48.4	93.6	39.6	6.1	50.1	50.7	59.4	70.0	56.3	87.4	77.1	49.3
RandLA-Net^[25]	77.8	73.1	69.9	57.7	82.9	44.6	73.6	47.7	52.3	94.5	45.4	26.9	48.4	74.9	61.8	73.8	59.9	82.7	79.2	62.1
KPConv^[38]	84.7	75.8	78.4	64.7	81.4	47.3	77.2	60.5	59.4	93.5	45.0	18.1	58.7	80.5	69.0	78.5	61.4	88.2	81.9	63.2
SparseConvNet^[39]	64.7	82.1	84.6	72.1	86.9	53.3	75.4	60.3	61.4	95.5	57.2	32.5	71.0	87.0	72.4	82.3	62.8	93.4	86.5	68.3
MinkowskiNet^[40]	85.9	81.8	83.2	70.9	84.0	52.1	85.3	66.0	64.3	95.1	54.4	28.6	73.1	89.3	67.5	77.2	68.3	87.4	85.2	72.7
PTv3^[30]	75.1	80.9	82.7	71.4	84.7	54.2	87.1	59.3	68.7	96.0	53.9	35.0	76.7	65.6	76.7	78.0	63.5	93.8	88.2	79.5
PTv3+MS-SFA	73.6	81.6	84.5	73.6	83.6	52.6	90.4	65.9	66.7	95.7	53.7	33.2	77.4	68.4	76.7	79.6	70.1	89.7	87.7	76.5

模型	mIoU	IoU
模型	mIoU	汽车	自行车	摩托车	卡车	其他车辆	行人	骑自行车	骑摩托车	道路	停车区	人行道	其他路面	建筑物	围栏	植被	树干	地形	杆子	标志
PTv3^[30]	65.5	96.6	52.6	71.5	85.3	63.5	77.5	88.6	0	95.3	59.2	84.3	0.5	79.4	57.1	84.9	65.8	73.0	64.1	45.3
PTv3+MS-SFA	66.7	96.0	53.7	78.0	92.1	59.0	78.1	89.2	0.1	95.1	59.4	84.6	0.1	80.0	59.2	84.5	68.4	72.0	64.6	48.1

方法	方法类别	mIoU/（%）
PointNet++^[8]	点	20.1
TangentConv^[41]	点	40.9
RandLA-Net^[25]	点	55.9
KPConv^[38]	点	58.8
SPVNAS^[42]	体素	67.0
PTv3^[30]	点	66.6
PTv3+MS-SFA	点+超体素	67.7