Bi3D++: Hybrid Bi-Domain Active Learning for Cross-Domain 3D Object Detection

Area of Science:

• Computer Vision
• Machine Learning
• Autonomous Systems

Background:

• Domain adaptation is crucial for 3D object detection, especially in autonomous driving where sensor and environmental variations cause domain discrepancies.
• Unsupervised domain adaptation (UDA) methods have improved performance but still lag behind fully supervised models due to significant domain gaps.
• Existing UDA methods struggle with the complexities of real-world autonomous driving scenarios, including sensor differences and environmental changes.

Purpose of the Study:

• To introduce Active Domain Adaptation (ADA) for 3D object detection, leveraging specific characteristics of autonomous driving like similar scenes and class imbalance.
• To propose a novel hybrid bi-domain active learning strategy, Bi3D++, for efficient knowledge transfer and performance enhancement in the target domain.
• To address the limitations of UDA by actively selecting and annotating crucial target data to bridge domain discrepancies.

Main Methods:

• Developed Bi3D++, a hybrid bi-domain active learning strategy for 3D object detection.
• Implemented target-like source data sampling based on scene-level and instance-level similarity to filter irrelevant source data.
• Introduced a hybrid active target sampling strategy considering rare-class similarity, intra-frame, and inter-frame diversity for comprehensive data selection.

Main Results:

• Bi3D++ outperforms state-of-the-art Unsupervised Domain Adaptation (UDA) methods in cross-domain 3D object detection tasks.
• Achieved superior performance with only 1% labeled target data compared to existing UDA approaches.
• Demonstrated consistent performance improvement with an increasing amount of annotated target data.

Conclusions:

• Active Domain Adaptation (ADA) is a promising direction for improving 3D object detection in autonomous driving.
• The proposed Bi3D++ strategy effectively addresses domain discrepancies by intelligently selecting informative data from both source and target domains.
• Bi3D++ offers a practical solution for enhancing 3D object detection models with minimal annotation effort, especially in challenging autonomous driving conditions.