Cross-attention guided discriminative feature selection for robust point cloud domain generalization

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Deep learning networks are prevalent for point cloud classification but struggle with domain discrepancies, leading to prediction errors.
• Domain generalization (DG) seeks to improve model performance on unseen data without retraining.
• Existing DG methods often overlook discriminative feature selection, a critical factor for robust generalization.

Purpose of the Study:

• To propose a novel domain generalization method for 3D point clouds that emphasizes discriminative feature selection.
• To improve the generalization performance of point cloud classification by effectively transferring contextual information.

Main Methods:

• Projecting point clouds into multiple views and utilizing a 2D adaptive feature extractor for semantic feature capture.
• Employing the DGCNN network to extract 3D spatial geometric features.
• Integrating an attention mechanism to fuse 2D semantic and 3D geometric features for discriminative feature selection.

Main Results:

• The proposed method demonstrates superior generalization performance compared to state-of-the-art techniques.
• Outperformed existing methods in both multi-source and single-source domain generalization tasks.
• Validated the effectiveness of discriminative feature selection for robust point cloud classification.

Conclusions:

• The novel approach successfully addresses the limitations of current domain generalization methods for point clouds.
• The fusion of multi-view semantic and geometric features via attention is key to achieving robust generalization.
• This work provides a significant advancement in achieving reliable 3D point cloud classification across diverse scenarios.