Open-CRB: Toward Open World Active Learning for 3D Object Detection.

Area of Science:

• Computer Vision
• Machine Learning
• Robotics

Background:

• Active learning (AL) enhances LiDAR-based 3D object detection using selected point clouds.
• Current AL methods do not address novel object detection in real-world streaming data.

Purpose of the Study:

• Investigate Open World Active Learning for 3D Object Detection (OWAL-3D).
• Develop a strategy to acquire informative point clouds containing new concepts.
• Address the challenge of identifying unknown objects in dynamic environments.

Main Methods:

• Propose Open Label Conciseness (OLC) for mining novel 3D objects efficiently.
• Integrate OLC with generic AL policies for OWAL-3D.
• Introduce the Open-CRB framework combining OLC and the CRB active learning method.
• Develop a comprehensive codebase supporting multiple baseline methods, 3D detectors, and datasets.

Main Results:

• OLC successfully adapts 3D detection models to open-world scenarios in a single selection round.
• The Open-CRB framework demonstrates superiority in recognizing both novel and known classes.
• Achieved efficient identification of new concepts with minimal annotation costs.

Conclusions:

• The proposed OLC strategy is effective for open-world 3D object detection.
• Open-CRB offers a flexible and superior approach for recognizing novel objects in real-world applications.
• The developed codebase facilitates future research in AL for 3D object detection.