MSSP: Multi-Scale Spatially-Constrained Partition for Unsupervised Semantic Segmentation of 3D Point Clouds
Authors: Zhenghao Zhang, Xinjie Wang, Wei Wang, Jun Zhang, Hanyun Wang
Organizations: College of Electronic Science and Technology, National University of Defense Technology, Changsha, China · School of Electronics and Communication Engineering, Sun Yat-sen University, Shenzhen, China
3D point cloud semantic segmentation is essential for real-world spatial understanding, yet the prohibitive cost of human annotations motivates unsupervised approaches that require no labels. Existing superpoint-based methods typically rely on spectral analysis at a fixed granularity, failing to capture the hierarchical semantic structures inherent in complex indoor scenes. To bridge this gap, we present a Multi-Scale Spatially-Constrained Partition (MSSP) framework that combines multi-scale spectral analysis with spatially-constrained clustering. Multi-scale spectral analysis constructs enriched superpoint descriptors across multiple clustering granularities; however, the resulting high-dimensional feature space calls for a structural prior to translate into cleaner segmentation. Spatially-constrained clustering supplies this prior by restricting superpoint merging to physically adjacent regions, imposing the spatial coherence needed for multi-scale features to be effective. Extensive experiments on S3DIS and ScanNet show that MSSP achieves the best mIoU among unsupervised methods on the main benchmarks, with particularly significant gains on S3DIS. Notably, our ablation reveals a regularize-then-enrich interaction: multi-scale features alone do not improve final segmentation, yet become highly effective when coupled with spatial regularization, underscoring that spatial coherence is aprerequisite for multi-scale representations in superpoint clustering.
Figures & tables
Figure 1: Overview of the MSSP framework. DINOv2 features are projected to 3D and distilled into a sparse 3D backbone; the resulting per-superpoint features hsm fork into two branches—multi-scale fusion feeds spectral primitive discovery, while spatially-constrained clustering regularizes the superpoint grouping—that converge at primitive assignment to yield pseudo-labels.
Figure 2: Qualitative comparison on ScanNet (top row) and S3DIS (bottom row). From left to right: representative baseline methods, MSSP (Ours), and Ground Truth. MSSP produces more coherent and accurate segments, particularly on structurally complex classes and at object boundaries.
Val Split
Online
Method
OA(%)
mAcc(%)
mIoU(%)
mIoU(%)
Supervised
PointNet++ [ 20 ]
—
33.9
PointCNN [ 14 ]
—
45.8
SparseConv [ 7 ]
—
72.5
Unsupervised
Table 1: Results on ScanNet dataset. Best unsupervised results in bold . Supervised methods report only online mIoU; “—” = metric not reported. † GrowSP does not use pretrained 2D visual features; MSSP, GrowSP++, and the other unsupervised baselines use pretrained DINO/DINOv2 features, so gains over GrowSP partly reflect stronger 2D pretraining.
Area 5
6-fold cross-validation
Method
OA(%)
mAcc(%)
mIoU(%)
OA(%)
mAcc(%)
mIoU(%)
Supervised
PointNet [ 19 ]
77.5
59.1
44.6
75.9
67.1
49.4
PointNet++ [ 20 ]
77.5
62.6
50.1
77.1
74.1
55.1
SparseConv [ 7 ]
88.4
69.2
60.8
89.4
78.1
69.2
Unsupervised
Table 2: Results on S3DIS dataset. Left: standard Area 5 protocol (train on Areas 1, 2, 3, 4, 6; test on Area 5). Right: 6-fold cross-validation (leave-one-area-out, averaged over Areas 1–6). Best unsupervised results in bold . † Does not use pretrained 2D visual features; MSSP and the other unsupervised baselines use pretrained DINO/DINOv2 features.
test on →
Area 1
Area 2
Area 3
Area 4
Area 5
Area 6
Mean
IIC [ 9 ]
3.7
3.8
3.8
4.0
3.8
3.7
3.8
PiCIE [ 4 ]
13.5
12.7
13.4
12.8
11.3
13.1
12.8
GrowSP [ 27 ]
24.2
21.9
26.1
25.0
23.7
27.9
24.8
PointDC [ 3 ]
23.6
20.9
24.6
19.5
20.1
29.7
23.1
PointDC-DINOv2 [ 3 ]
33.8
29.6
33.8
31.7
32.2
36.9
33.0
LogoSP [ 26 ]
43.8
37.5
47.0
40.7
44.9
47.9
43.6
Table 3: Cross-dataset generalization (mIoU%). Best unsupervised results in bold .
Setting
SC
MS
OA(%)
mAcc(%)
mIoU(%)
LogoSP baseline
82.8
55.9
46.5
+SC only
✓
83.2
51.7
45.3
+MS only
✓
80.8
57.4
45.3
+SC + MS (Ours)
✓
✓
84.2
56.8
49.2
Table 4: Ablation study on S3DIS Area 5. SC: Spatially-constrained clustering. MS: Multi-scale spectral analysis. Best in bold .
Figure 3: Ablation visualization on S3DIS. Neither MS nor SC alone improves over the baseline, but their combination produces markedly better segmentation.
K
Dim
mIoU(%)
mAcc(%)
{80}
384
33.2
46.9
{40}
384
34.1
49.9
{20}
384
31.4
46.3
{80,40}
768
34.4
50.5
{40,20}
768
34.7
48.6
{80,20}
768
34.8
49.8
Table 5: Multi-scale granularity ablation on ScanNet. All variants use the same experimental setup, varying only the set of KMeans scales K . Best in bold .
Key Laboratory of Big Data & Artificial Intelligence in Transportation (Beijing Jiaotong University), Ministry of Education · Frontiers Science Center for Smart High-speed Railway System, Beijing Jiaotong University
College of Systems Engineering, National University of Defense Technology · Laboratory for Big Data and Decision, National University of Defense Technology