MA-ST3D: Motion Associated Self-Training for Unsupervised Domain Adaptation on 3D Object Detection

Area of Science:

• Computer Vision
• Machine Learning
• Robotics

Background:

• Unsupervised domain adaptation (UDA) is crucial for 3D object detection to reduce annotation costs.
• Self-training (ST) is a common UDA technique, but suffers from imprecise pseudo-labels due to domain shift.
• Accurate pseudo-labels are essential for effective self-training in 3D object detection.

Purpose of the Study:

• To develop a novel self-training unsupervised domain adaptation (ST-UDA) framework for 3D object detection.
• To generate high-quality pseudo-labels by leveraging spatial and temporal consistency in 3D point cloud sequences.
• To improve the performance of 3D object detectors in domain adaptation scenarios.

Main Methods:

• Introduced motion-associated self-training for 3D object detection (MA-ST3D).
• Employed a global-local pathway (GLP) architecture utilizing intra-frame and inter-frame consistencies.
• Integrated global and local memory modules to stabilize pseudo-labels and a motion-aware loss function.

Main Results:

• MA-ST3D achieved state-of-the-art (SOTA) performance across all evaluated unsupervised domain adaptation settings.
• The method demonstrated superior performance compared to weakly supervised domain adaptation on Kitti and NuScenes benchmarks.
• Evaluated on Waymo, Kitti, and nuScenes datasets, confirming robust domain adaptation capabilities.

Conclusions:

• MA-ST3D effectively generates high-quality pseudo-labels for 3D object detection in unsupervised domain adaptation.
• The proposed framework significantly enhances 3D object detection performance under domain shift.
• MA-ST3D offers a promising solution for reducing annotation costs in 3D object detection.