ST3D++: Denoised Self-Training for Unsupervised Domain Adaptation on 3D Object Detection.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Unsupervised domain adaptation is crucial for 3D object detection, enabling models to generalize to new environments without labeled data.
• Existing methods often struggle with noisy pseudo-labels generated during self-training, hindering performance.
• Addressing pseudo-label noise is essential for effective domain adaptation in 3D object detection.

Purpose of the Study:

• To propose ST3D++, a novel self-training method for unsupervised domain adaptation in 3D object detection.
• To develop a holistic pseudo-label denoising pipeline to mitigate noise in pseudo-label generation and training.
• To enhance the quality and stability of pseudo-labels and improve model robustness against noisy data.

Main Methods:

• ST3D++ employs a pre-training strategy with random object scaling (ROS) to reduce scale bias.
• A hybrid quality-aware triplet memory is used for improved pseudo-label generation.
• A source data assisted training strategy and curriculum data augmentation are utilized to refine training signals and prevent over-fitting.

Main Results:

• ST3D++ achieves state-of-the-art performance on four benchmark datasets (Waymo, KITTI, Lyft, nuScenes) for car, pedestrian, and bicycle detection.
• The method significantly outperforms baselines, with improvements ranging from 9.6% to 38.16% on Waymo → KITTI.
• ST3D++ surpasses fully supervised results on the KITTI 3D object detection benchmark when target prior is available.

Conclusions:

• ST3D++ effectively adapts 3D object detectors to new domains without annotations by meticulously refining pseudo-labeled data and denoising training signals.
• The proposed denoising pipeline and training strategies are key to achieving superior performance in unsupervised domain adaptation.
• The method demonstrates the potential for high-accuracy 3D object detection in challenging, unlabeled target domains.