Continual Learning Strategy in One-Stage Object Detection Framework Based on Experience Replay for Autonomous Driving...

Area of Science:

• Computer Vision
• Machine Learning
• Autonomous Driving Systems

Background:

• Object detection is crucial for autonomous driving vehicles (ADV).
• ADV deployment growth necessitates detecting an expanding variety of objects.
• Existing models struggle to adapt to new classes without performance degradation.

Purpose of the Study:

• To propose a novel continual learning method for object detection.
• To enable models to learn new object classes on the fly.
• To mitigate catastrophic forgetting in continual learning scenarios.

Main Methods:

• A novel continual learning method for object detection is proposed.
• The method learns new object classes alongside prior knowledge.
• It incorporates a cumulative memory strategy to retain previous learning.

Main Results:

• The proposed ER method demonstrated a low mean Average Precision (mAP) drop of 4.3% on PASCAL VOC 2007.
• This performance is competitive compared to all-classes learning.
• The method achieved the lowest performance drop among existing prior art.

Conclusions:

• The proposed continual learning method effectively addresses the challenge of expanding object detection classes in ADV.
• It offers a robust solution to prevent catastrophic forgetting.
• The method shows significant promise for enhancing the adaptability of object detection models for autonomous driving.