PointCAT: Contrastive Adversarial Training for Robust Point Cloud Recognition.

Area of Science:

• Computer Vision
• Machine Learning
• 3D Data Processing

Background:

• Point cloud recognition models excel in many applications but are vulnerable to natural corruptions and adversarial perturbations.
• Existing methods often struggle to generalize robustness across diverse noise types.

Purpose of the Study:

• To propose a novel method, Point-Cloud Contrastive Adversarial Training (PointCAT), for significantly boosting the general robustness of point cloud recognition.
• To address the limitations of logit-level constraints by introducing feature-level mechanisms.

Main Methods:

• PointCAT employs feature-level constraints to minimize the gap between clean and corrupted point cloud representations.
• Utilizes a supervised contrastive loss for hypersphere representation alignment and uniformity.
• Incorporates dynamic prototype-guided centralizing losses to maintain feature cluster integrity.
• Adversarially trains a noise generator concurrently with the recognition model, differing from prior inner-loop gradient-based attacks.

Main Results:

• PointCAT demonstrates superior performance compared to baseline methods across various corruption types.
• Achieves significant enhancements in robustness for diverse point cloud recognition models.
• Effectively handles isotropic point noises, LiDAR simulated noises, random point dropping, and adversarial perturbations.

Conclusions:

• PointCAT offers a robust and effective solution for improving the resilience of point cloud recognition systems.
• The feature-level contrastive and centralizing approach provides a more generalized robustness against various data corruptions.