Towards Robust Point Cloud Recognition With Sample-Adaptive Auto-Augmentation.

Area of Science:

• Computer Vision
• Machine Learning
• 3D Data Processing

Background:

• 3D perception requires robustness against data corruption.
• Conventional augmentation methods fail to adapt to sample structures, causing uneven enhancement.
• Existing methods lack sufficient real-world corrupted 3D point cloud data.

Purpose of the Study:

• To develop a sample-adaptive auto-augmentation framework for robust 3D perception.
• To improve the handling of corrupted 3D point cloud data by considering intrinsic sample structures.
• To introduce novel datasets for training and evaluating models on real-world corrupted data.

Main Methods:

• Proposed AdaptPoint++ auto-augmentation framework with imitator and discriminator.
• Imitator uses Position-aware Feature Extraction, Deformation Controller, and Mask Controller for adaptive corruption simulation.
• Introduced Structure Reconstruction-assisted learning and perception-guidance feedback.
• Created two new datasets: ScanObjectNN-C and MVPNET-C.

Main Results:

• AdaptPoint++ achieves state-of-the-art performance on multiple corruption benchmarks.
• The framework effectively generates corrupted samples tailored to intrinsic data structures.
• Structure Reconstruction-assisted learning enhances classifier robustness against corruption.

Conclusions:

• Sample-adaptive augmentation is superior to conventional methods for 3D corruption robustness.
• AdaptPoint++ provides an effective solution for enhancing 3D perception in corrupted environments.
• The novel datasets facilitate further research and development in corrupted 3D data processing.