Parameter-Efficient Fine-Tuning in Spectral Domain for Point Cloud Learning.

Area of Science:

• Computer Vision
• Machine Learning
• 3D Data Processing

Background:

• Pre-training enhances point cloud models but requires computationally expensive full fine-tuning.
• Existing methods are storage-intensive and demand significant computational resources for downstream tasks.

Purpose of the Study:

• To introduce a novel Parameter-Efficient Fine-Tuning (PEFT) method for point cloud models.
• To address the storage and computational challenges associated with full fine-tuning.
• To improve the efficiency of transferring general knowledge to downstream point cloud tasks.

Main Methods:

• Propose PointGST (Point cloud Graph Spectral Tuning), a PEFT method that freezes pre-trained models.
• Introduce a lightweight Point Cloud Spectral Adapter (PCSA) for spectral domain fine-tuning.
• Transfer point tokens to the spectral domain to de-correlate spatial confusion and incorporate intrinsic task-specific information.

Main Results:

• PointGST outperforms full fine-tuning on challenging point cloud datasets.
• Achieves superior accuracies: 99.48% (ScanObjNN OBJ_BG), 97.76% (OBJ_ONLY), 96.18% (PB_T50_RS).
• Reduces trainable parameters to only 0.67% of the total, establishing a new state-of-the-art.

Conclusions:

• PointGST offers an efficient solution for point cloud learning by minimizing training costs.
• The spectral domain adaptation effectively transfers general knowledge for downstream tasks.
• Achieves state-of-the-art performance with significantly reduced parameter requirements.