Reproduce “MapFM: Foundation Model-Driven HD Mapping with Multi-Task Contextual Learning” 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
Leonid Ivanov, Vasily Yuryev, Dmitry Yudin
MapFM enhances online vectorized High-Definition (HD) map generation for autonomous driving by integrating large-scale vision foundation models (DINOv2) and a multi-task learning approach. This method achieves state-of-the-art performance with an mAP of 69.0% on the nuScenes dataset, representing an improvement of up to 2.9% mAP over previous baselines.
Key insights
- Integrating powerful foundation models like DINOv2 as the feature extractor significantly boosts accuracy in online vectorized HD map generation, especially when fine-tuned.
- Multi-task learning, particularly through auxiliary BEV semantic segmentation, enriches the model's contextual understanding and improves the primary vectorized map prediction task.
- Targeted fine-tuning of the foundation model's last layers is crucial to unlock its full potential for specific downstream tasks, outperforming frozen features.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.