Structure-Aware Domain Adaptive Point Mamba for Point Cloud Completion

Point cloud completion (PCC) aims to recover complete 3D shapes from partial observations, yet fully supervised models often degrade substantially when transferred across domains. Existing unsupervised domain adaptation (UDA) methods alleviate this issue through reconstruction, self-supervision, distillation, or feature alignment, but generally lack a unified structural perspective. Their patch-wise, structure-agnostic alignment may associate regions with different structural roles, causing unreliable correspondences and structural drift, while existing Euclidean alignment methods may be limited in modeling multi-level semantic relations across domains. To address these limitations, we propose StructDAMamba, a structure-aware domain adaptive Point Mamba framework that jointly models intra-domain patch dependencies and cross-domain semantic hierarchies. Specifically, Global Structure-Aware Spatial Alignment (GSSA) captures local-to-global patch relationships and aligns compact structural representations at the distribution level, reducing noise-induced mismatches without relying on explicit patch correspondences. Hyperbolic Domain Feature Aggregation (HDFA) projects source- and target-domain features into hyperbolic space for intra-domain integration and cross-domain aggregation, encouraging more transferable representations. Extensive experiments across synthetic and real-world benchmarks demonstrate that StructDAMamba consistently achieves state-of-the-art completion performance while maintaining high computational efficiency.