Parameter-Efffcient Topology-Guided Cross-Scale Adapter for Point Cloud Learning

Recently, large-scale pre-training has become a dominant paradigm for improving point cloud representations and enabling strong transfer to downstream three-dimensional (3D) tasks. However, adapting large pre-trained point-cloud transformers in practice still often relies on full fine-tuning, which is storage-intensive and computationally demanding when multiple tasks or domains must be supported. Moreover, for real scans, the main obstacle is not only the parameter budget but also the topology shift induced by density variation, occlusion, missing regions, and background clutter, which corrupts local neighborhoods and makes token-level adaptation unstable. To address these issues, we pro-pose a novel parameter-efficient fine-tuning (PEFT) framework for point clouds, called TGCS (Topology-Guided Cross-Scale adapter). TGCS freezes the pre-trained backbone and introduces a lightweight, trainable tuning branch that performs topology-conditioned residual calibration across transformer blocks. The core idea is built on two observations: (1) under a frozen backbone, feature-space prompts and adapters may be misled by unreliable semantic tokens when neighborhood topology is distorted, and (2) topology corruption is inherently multi-scale, so effective tuning should couple explicit topology cues with cross-scale context. Concretely, TGCS combines Cross-Scale Token Mixing (CS-Mixing), Saliency-Aware Token Gating (SA-Gating), and a Topology-Guided Cross-Scale Adapter (TG-Adapter) that conditions residual updates on multi-scale topology descriptors computed from token anchors, including density and dispersion statistics as well as eigenvalue-derived local shape cues. Extensive experiments on ScanObjectNN, ModelNet40, and ShapeNetPart demonstrate that TGCS consistently improves the accuracy-efficiency trade-off across MAE-style and GPT-style backbones. Notably, with Point-MAE, TGCS tunes only 0.6M parameters (2.68%) yet improves the hardest ScanObjectNN setting PB_T50 RS from 85.18% to 88.03%. With the stronger PointGPT-L back-bone, TGCS achieves 98.97%, 97.42%, and 95.00% on 0BJ_BG, 0BJ_ONLY, and PB_T50_RS, respectively while tuning only 2.2M parameters, establishing the state-of-the-art performance under an efficient fine-tuning regime. TGCS also yields stable gains in few-shot classification and preserves competitive part-segmentation mIoU with a compact tunable budget, validating topology-guided cross-scale conditioning as a practical solution for resource-efficient point cloud adaptation.