TopAdapter: Topology-Aware Prompt Tuning for Efficient Point Cloud Understanding

Point cloud data, with its inherent geometric and topological structures, plays a critical role in 3D vision tasks. However, existing parameter-efficient fine-tuning (PEFT) methods predominantly focus on input token prompting, overlooking the intrinsic geometric information. To address this limitation, we propose TopAdapter, a novel PEFT framework that enhances geometric perception by injecting local topological information into pre-trained 3D vision models. TopAdapter leverages 0D, 1D, and 2D simplices from algebraic topology as fundamental building blocks, introducing two core modules: the Topology Injection module (ToInjection) and the Topology Transfer module (ToTransfer). ToInjection constructs multi-scale topological features using a simplex generator and dynamically fuses them with semantic features via a geometric controller, thereby enhancing geometric adaptability. ToTransfer propagates these topological primitives across Transformer layers, ensuring efficient transmission of geometric information. Extensive experiments demonstrate that TopAdapter outperforms existing PEFT methods, achieving performance comparable to full fine-tuning across various benchmarks.

3D point clouds, like those captured by self-driving cars or robots, are collections of points in space. Existing efficient fine-tuning methods for point cloud models mainly focus on individual points, ignoring the connections between them—such as how points form edges, surfaces, and shapes. We propose TopAdapter, a new method that explicitly captures these geometric relationships using simple building blocks: points (0D), edges (1D), and triangles (2D). By injecting this topological information into pre-trained models and propagating it through all layers, TopAdapter helps models better understand 3D structure. Remarkably, our method uses only 0.5 million trainable parameters (less than 1% of the full model) while achieving state-of-the-art results on standard benchmarks, often matching or exceeding full fine-tuning. This makes advanced 3D vision more accessible and environmentally sustainable.