Parameter Efficient Point Cloud Prompt Tuning for Unified Point Cloud Understanding

The prevailing paradigm in 3D vision involves fully fine-tuning all the backbone parameters of pre-trained models. However, this approach poses challenges due to the large number of parameters requiring tuning, resulting in unexpected storage demands. To address these issues and alleviate the computational cost and storage burden associated with full fine-tuning, we propose Point Cloud Prompt Tuning (PCPT) as an effective method for large Transformer models in point cloud processing. PCPT offers a powerful and efficient solution to mitigate the costs associated with full fine-tuning. Drawing inspiration from recent advancements in efficient tuning of large-scale language models and 2D vision models, PCPT leverages less than 0.05 $\%$ of trainable parameters, while keeping the pre-trained parameters of the Transformer backbone unchanged. To evaluate the effectiveness of PCPT, extensive experiments were conducted on four discriminative datasets (ModelNet40, few-shot ModelNet40, ScanObjectNN, ShapeNetPart) and four generation datasets (PCN, MVP, ShapeNet55, and ShapeNet34/Unseen21). The results demonstrate that the task-specific prompts utilized in PCPT enable the Transformer model to adapt effectively to the target domains, yielding results comparable to those obtained through other full fine-tuning methods. This highlights the versatility of PCPT across various domains and tasks.