Agentic data science with the Data Agent Kit
cloud.google.com·7h ago
AI Research
arXiv CS.AI·Aug 15·5 min read
PseudoMapLabeler: Confidence-Aware Pseudo-Label Generation for Semi-Supervised Online Mapping
TL;DR
PseudoMapLabeler introduces a semi-supervised learning framework that generates reliable pseudo-labels for online HD map construction, improving model performance by 6.1 mAP with limited labeled data.
✦ Why It Matters
Implement confidence-aware pseudo-labeling to enhance your model's performance with limited labeled data.
Key Takeaways
- Utilize a teacher-student framework for model training.
- Apply confidence maps to assess prediction reliability.
- Implement spatial clipping to refine map elements.
- Leverage pseudo-labels to train models with minimal labeled data.
How It Works
The PseudoMapLabeler framework begins with a teacher model trained on limited labeled data. It generates confidence maps using a Beta distribution to evaluate the reliability of map predictions over time.
Instead of discarding all uncertain predictions, a spatial clipping technique is employed to retain only high-confidence areas, refining the map elements. These refined elements serve as improved map priors, enhancing the teacher model's predictions on unlabeled data, which are then used as pseudo-labels for training a student model.