MapRF uses NeRF-guided self-training for weakly supervised HD map construction

Researchers have developed MapRF, a novel framework for constructing high-definition (HD) maps for autonomous driving systems using only 2D image labels. This weakly supervised approach leverages Neural Radiance Fields (NeRF) to generate pseudo-labels, which are then used to iteratively refine the map network through self-training. A key component, Map-to-Ray Matching, helps align map predictions with camera rays to reduce errors. Experiments on the Argoverse 2 and nuScenes datasets show MapRF achieves performance comparable to fully supervised methods, making HD map construction more scalable and cost-effective. AI

Summary written by gemini-2.5-flash-lite from 1 source. How we write summaries →

IMPACT Enables more scalable and cost-effective HD map construction for autonomous driving by reducing reliance on costly 3D annotations.

RANK_REASON Academic paper detailing a new method for HD map construction. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CV →

• MapRF
• NeRF
• Argoverse 2
• nuScenes
• Hongyu Lyu

• paper
• other