SFD-ADNet: Spatial-Frequency Dual-Domain Adaptive Deformation for Point Cloud Data Augmentation.

Area of Science:

• Computer Vision
• Machine Learning
• 3D Data Processing

Background:

• Current 3D point cloud enhancement methods often fail to preserve global structure and adapt to diverse degradations due to predefined transformations.
• Existing techniques struggle with illogical deformations and limited adaptability.

Purpose of the Study:

• To propose SFD-ADNet, an adaptive deformation framework utilizing a dual spatial-frequency domain for 3D point cloud augmentation.
• To generate structurally aware and task-relevant augmented samples by learning deformation parameters.

Main Methods:

• Employs a dual spatial-frequency domain approach for adaptive deformation.
• Utilizes a hierarchical sequence encoder and Mamba-based predictor for spatial domain analysis.
• Incorporates a multi-scale dual-channel mechanism with adaptive Chebyshev polynomials for frequency domain analysis.

Main Results:

• SFD-ADNet reduces mCE metrics by over 20% for PointNet++ and other backbone networks on ModelNet40-C and ScanObjectNN-C datasets.
• Achieves state-of-the-art robustness while preserving critical geometric structures in 3D point clouds.
• Demonstrates consistently improved robustness against diverse point cloud attacks.

Conclusions:

• SFD-ADNet effectively enhances 3D point cloud robustness through adaptive space-frequency deformation.
• The framework offers a universal augmentation module adaptable to various point cloud processing tasks.
• Validates the efficacy of joint spatial and frequency domain modeling for robust 3D point cloud learning.