Source-Free Domain Adaptation for Geospatial Point Cloud Semantic Segmentation

TL;DR

LoGo tackles source-free unsupervised domain adaptation for semantic segmentation of heterogeneous geospatial point clouds by integrating a parameter-efficient mean-teacher framework with class-balanced local prototypes and a global optimal-transport distribution alignment, followed by a local-global dual-consensus filtering of pseudo-labels. The method addresses long-tailed class distributions and cross-sensor/cross-scene domain gaps, achieving state-of-the-art results on STPLS3D→H3D and DALES→Toronto-3D benchmarks. Key contributions include intra-class anchor mining for robust prototypes, entropy-regularized OT for global class priors, and a dual-consensus pseudo-label selection to curb noise. The approach demonstrates practical, privacy-preserving adaptation without source data, with strong performance gains and insights into hyperparameter robustness and ablations, supporting scalable geospatial mapping in real-world remote sensing workflows.

Abstract

Semantic segmentation of 3D geospatial point clouds is pivotal for remote sensing applications. However, variations in geographic patterns across regions and data acquisition strategies induce significant domain shifts, severely degrading the performance of deployed models. Existing domain adaptation methods typically rely on access to source-domain data. However, this requirement is rarely met due to data privacy concerns, regulatory policies, and data transmission limitations. This motivates the largely underexplored setting of source-free unsupervised domain adaptation (SFUDA), where only a pretrained model and unlabeled target-domain data are available. In this paper, we propose LoGo (Local-Global Dual-Consensus), a novel SFUDA framework specifically designed for geospatial point clouds. At the local level, we introduce a class-balanced prototype estimation module that abandons conventional global threshold filtering in favor of an intra-class independent anchor mining strategy. This ensures that robust feature prototypes can be generated even for sample-scarce tail classes, effectively mitigating the feature collapse caused by long-tailed distributions. At the global level, we introduce an optimal transport-based global distribution alignment module that formulates pseudo-label assignment as a global optimization problem. By enforcing global distribution constraints, this module effectively corrects the over-dominance of head classes inherent in local greedy assignments, preventing model predictions from being severely biased towards majority classes. Finally, we propose a dual-consistency pseudo-label filtering mechanism. This strategy retains only high-confidence pseudo-labels where local multi-augmented ensemble predictions align with global optimal transport assignments for self-training.

Figures (9)

• Figure 1: Overview of domain adaptation paradigms. Model Input, Model Output. The snowflake indicates frozen parameters, while the flame indicates trainable parameters. (a) No Adaptation; (b) Conventional Domain Adaptation; (c) Source-Free Unsupervised Domain Adaptation.
• Figure 2: Overview of the proposed LoGo framework. The architecture adheres to a parameter-efficient Mean-Teacher paradigm, where only the BN layers are learnable (indicated by the flame icon), while other parameters remain frozen (snowflake icon). The adaptation process begins by aggregating features from multi-augmented inputs via the Class-Balanced Local Prototype Estimation (CBLPE) module, which employs an intra-class anchor mining strategy to estimate robust class prototypes. Subsequently, the Global Distribution Alignment (GDA) module solves an Optimal Transport problem to generate a global assignment plan $\mathbf{Q}^*$ by minimizing transport costs under class prior constraints. Finally, the Local-Global Dual-Consensus Filtering (LGDCF) mechanism selects reliable pseudo-labels by identifying the intersection ($\cap$) between local classifier predictions and global OT assignments. The student model is then supervised by these refined labels on strongly augmented data via Cross-Entropy (CE) loss, while the teacher model evolves through Exponential Moving Average (EMA).
• Figure 3: Visualization of the domain adaptation scenario from photogrammetry-derived point clouds (STPLS3D) to UAV-based LiDAR point clouds (H3D). The top row compares the representative scenes, while the bottom row displays their corresponding semantic label spaces.
• Figure 4: Visualization of the domain adaptation scenario from ALS point clouds (DALES) to MLS point clouds (T3D). The top row compares the representative scenes, while the bottom row displays their corresponding semantic label spaces.
• Figure 5: Qualitative comparison of global semantic segmentation results on the H3D dataset. From left to right: Source-only, SHOT, TTYD, LoGo, and Ground Truth.