Reproduce “ST3D: Self-training for Unsupervised Domain Adaptation on 3D Object Detection” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jihan Yang, Shaoshuai Shi, Zhe Wang, Hongsheng Li, Xiaojuan Qi
ST3D presents a self-training pipeline for unsupervised domain adaptation on 3D object detection, enabling models to adapt to new, unlabeled environments. The framework achieves state-of-the-art performance on various autonomous driving datasets, boosting AP3D by up to 74.30% over source-only baselines and occasionally outperforming fully supervised models.
Key insights
- Relying solely on classification confidence for 3D pseudo-label selection is insufficient; an IoU-based quality-aware criterion is essential for accurately assessing localization precision.
- Self-training stability and pseudo-label quality are enhanced by a "triplet box partition" that filters ambiguous predictions and by a "memory ensemble and voting" mechanism that integrates historical pseudo-labels.
- Curriculum Data Augmentation, by gradually increasing augmentation intensity, prevents overfitting to "easy" pseudo-labeled examples and fosters learning from more diverse and challenging scenarios.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.