Submitted by Zhang Zhenghao.

Submission data

Full name	MSSP: Multi-Scale Spatial Partition for Unsupervised 3D Semantic Segmentation
Description	MSSP (Multi-Scale Spatial Partition) is an unsupervised 3D semantic segmentation method built upon LogoSP (CVPR 2025). It follows a two-stage pipeline: Stage 1 distills DINOv2 features from multi-view images into a 3D sparse convolutional backbone (Res16FPN18) via cosine similarity loss. Stage 2 performs spectral analysis on a superpoint affinity graph to discover semantic primitives. MSSP introduces three innovations: (1) multi-scale spectral analysis that constructs superpoint descriptors at multiple clustering granularities (K={80,40,20}) to capture semantic patterns at varying spatial extents; (2) per-scene stratified sampling that allocates the spectral budget equally across all training scenes for balanced eigendecomposition; (3) diversity loss based on entropy maximization to mitigate minority class collapse. The model is trained on ScanNet training set without any labeled data.
Input Data Types	Uses Color,Uses Geometry        Uses 2D,Uses 3D
Programming language(s)	Python with CUDA
Hardware	Intel Core i9-14900K,NVIDIA GeForce RTX 4090
Submission creation date	1 Jun, 2026
Last edited	1 Jun, 2026
Last uploaded	1 Jun, 2026

3D semantic label results