[论文解读] P-SLCR: Unsupervised Point Cloud Semantic Segmentation via Prototypes Structure Learning and Consistent Reasoning

Lixin Zhan, Jie Jiang+4 more|arXiv (Cornell University)|Mar 6, 2026

3D Shape Modeling and Analysis被引用 0

一句话总结

P-SLCR 在无监督点云语义分割方面提出了一种双原型库（一致性与模糊性）结合结构学习与一致推理的框架，在 S3DIS、SemanticKITTI 和 ScanNet 上达到无监督方法的最新水平。

ABSTRACT

Current semantic segmentation approaches for point cloud scenes heavily rely on manual labeling, while research on unsupervised semantic segmentation methods specifically for raw point clouds is still in its early stages. Unsupervised point cloud learning poses significant challenges due to the absence of annotation information and the lack of pre-training. The development of effective strategies is crucial in this context. In this paper, we propose a novel prototype library-driven unsupervised point cloud semantic segmentation strategy that utilizes Structure Learning and Consistent Reasoning (P-SLCR). First, we propose a Consistent Structure Learning to establish structural feature learning between consistent points and the library of consistent prototypes by selecting high-quality features. Second, we propose a Semantic Relation Consistent Reasoning that constructs a prototype inter-relation matrix between consistent and ambiguous prototype libraries separately. This process ensures the preservation of semantic consistency by imposing constraints on consistent and ambiguous prototype libraries through the prototype inter-relation matrix. Finally, our method was extensively evaluated on the S3DIS, SemanticKITTI, and Scannet datasets, achieving the best performance compared to unsupervised methods. Specifically, the mIoU of 47.1% is achieved for Area-5 of the S3DIS dataset, surpassing the classical fully supervised method PointNet by 2.5%.

研究动机与目标

• 为原始三维点云的无标注语义分割提供动机。
• 引入双原型库（一致性 vs 模糊性）以通过结构学习与推理引导学习。
• 开发机制以选择高质量的一致点并细化模糊点。
• 通过原型互相关矩阵在原型库之间强制语义一致性。
• 在室内与室外数据集上证明相对于无监督基线的最先进性能。

提出的方法

• 构建两个原型库（一致性与模糊性），通过对局部聚类中心的 EMA 更新。
• 通过网络预测与伪标签的一致性阈值定义一致点；将数据分成一致集和模糊集。
• 计算一致结构学习损失以使一致点特征与一致原型对齐。
• 形成语义关系一致推理约束，使用一致原型与模糊原型的独立相似度矩阵并最小化它们的分布差。
• 在 L_total 中将基于交叉熵的监督信号与 L_sl（structureLearning）和 L_cr（consistent reasoning）损失结合来训练模型。

Figure 1: P-SLCR architecture for structural learning and consistent reasoning through a learnable prototype-based library. The consistent points are continuously expanding and increasingly resembling the prototype.

实验结果

研究问题

• RQ1 双原型库（一致性 vs 模糊性）能否实现对点云的有效无监督语义分割？
• RQ2 一致结构学习是否提高点特征与原型表示之间的对齐？
• RQ3 语义关系一致推理是否能在一致与模糊原型之间保持并传播语义一致性？

主要发现

Method	OA(%)	mAcc(%)	mIoU(%)
PointNet	77.5	59.1	44.6
PointNet++ (Supervised)	77.5	62.6	50.1
SparseConv	88.4	69.2	60.8
GrowSP	78.4	57.2	44.5
U3DS 3	75.5	55.8	42.8
PointDC	54.1	-	22.6
KMeans	22.1	21.2	9.4
DBSCAN	17.5	19.8	9.2
IIC-S-PFH	31.2	16.3	9.1
PICIE-S-PFH	-	-	-
P-SLCR (ours)	79.7	57.2	47.1

• P-SLCR 在无监督设定下获得 S3DIS Area-5 的 mIoU 47.1%，超越经典全监督 PointNet 2.5% 的 mIoU。
• 在 S3DIS 6-Fold 上，P-SLCR 达到 47.5% 的 mIoU，多个指标上超越 GrowSP。
• 在 SemanticKITTI（验证集）上，P-SLCR 达到 15.3% 的 mIoU 和 55.9% 的 OA，超过此前的无监督方法。
• 在 SemanticKITTI 测试集上，P-SLCR 达到 15.9% 的 mIoU 和 61.4% 的 OA，OA/mIoU 方面超越 GrowSP 与 U3DS3。
• 在 ScanNet 上，P-SLCR 达到 29.0% 的 mIoU 和 61.4% 的 OA，在若干指标上优于 GrowSP 与 U3DS3。
• 消融研究表明移除一致推理会显著下降 mIoU，证实其关键作用。

Figure 2: An overview of our P-SLCR with features extracted by SparseConv, pseudo-labels and predictions obtained by clustering and segmenting the header. The proposed reliability classifies the scene into consistent and ambiguous points and establishes an updatable prototype library to construct co