CO-Net++: A Cohesive Network for Multiple Point Cloud Tasks at Once With Two-Stage Feature Rectification.

Area of Science:

• Computer Vision
• Machine Learning
• 3D Data Analysis

Background:

• Optimizing multiple point cloud tasks across diverse datasets presents challenges due to domain heterogeneity and parameter entanglement.
• Existing methods struggle to effectively balance universal feature extraction with task-specific adaptations.

Purpose of the Study:

• To introduce CO-Net++, a novel framework for collective optimization of diverse point cloud tasks.
• To enhance performance in 3D object detection and semantic segmentation by addressing parameter entanglement.
• To enable robust incremental learning and prevent catastrophic forgetting in new point cloud tasks.

Main Methods:

• Development of CO-Net++, a cohesive framework with a two-stage feature rectification strategy (TFRS).
• Task-shared parameter optimization in the backbone using sign-based gradient surgery to handle domain conflicts.
• Task-specific parameter integration in the second stage after freezing shared parameters.

Main Results:

• CO-Net++ achieves exceptional performance in both 3D object detection and 3D semantic segmentation tasks.
• The framework demonstrates significant improvements by mitigating conflicting optimization from parameter entanglement.
• CO-Net++ exhibits strong incremental learning capabilities, preventing catastrophic amnesia on new tasks.

Conclusions:

• CO-Net++ offers an effective solution for multi-task point cloud learning across heterogeneous domains.
• The two-stage feature rectification strategy successfully distinguishes and optimizes universal and task-specific features.
• The proposed method generalizes well to new point cloud tasks, showcasing adaptability and robustness.