SDCoT++: Improved Static-Dynamic Co-Teaching for Class-Incremental 3D Object Detection.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Deep learning excels at 3D object detection but struggles with catastrophic forgetting during incremental learning.
• Catastrophic forgetting, where performance degrades on old classes when learning new ones, hinders real-world AI applications requiring continuous learning.
• Class co-occurrence in scenes exacerbates forgetting and model confusion in incremental 3D object detection.

Purpose of the Study:

• To propose a novel static-dynamic co-teaching framework to address catastrophic forgetting in incremental 3D object detection.
• To enhance the ability of AI systems to continuously learn new object categories without performance degradation on previously learned ones.
• To develop a robust method that mitigates confusion caused by frequent co-occurrences of old and new classes.

Main Methods:

• Introduced a static-dynamic co-teaching approach with a student model and two teachers: a static teacher for old knowledge and a dynamic teacher for new knowledge.
• Generated pseudo labels for old classes from both static and dynamic teachers to mitigate co-occurrence issues during incremental learning.
• Calibrated base class probabilities to balance class occurrences and improve pseudo-label selection, enhancing robustness against varying class frequencies.

Main Results:

• The static-dynamic co-teaching method significantly outperforms baseline approaches in incremental 3D object detection.
• The framework demonstrates superior performance in preserving knowledge of previously trained classes while learning new ones.
• Experiments validated the method's effectiveness across diverse indoor and outdoor benchmark datasets.

Conclusions:

• The proposed static-dynamic co-teaching approach effectively overcomes catastrophic forgetting in incremental 3D object detection.
• The framework's backbone-agnostic nature allows seamless integration with various 3D detection architectures like VoteNet, 3DETR, and CAGroup3D.
• This research advances continuous learning capabilities for AI in real-world 3D perception tasks.