Reproduce “ADMap: Anti-disturbance framework for reconstructing online vectorized HD map” on OpenResearch

About this paper

Haotian Hu, Fanyi Wang, Yaonong Wang, Laifeng Hu, Jingwei Xu, Zhiwang Zhang

The ADMap framework introduces an anti-disturbance approach for vectorized HD map construction, mitigating point sequence jitter to produce more stable and accurate map elements for autonomous driving. It reduced the Average Chamfer Distance between predicted and ground truth instances by 0.079 compared to MapTR, and improved multimodal mean Average Precision by 5.5% on nuScenes.

Key insights

• Mitigating point sequence jitter requires a combined approach that includes enhanced multi-scale feature perception, refined inter- and intra-instance interactions, and explicit geometric constraints in the loss function.
• A loss function that explicitly models vector direction differences and weights them based on the ground truth's topological complexity can effectively guide the model to focus on critical, less predictable map segments.
• Separating feature processing for training (providing multi-scale details) and inference (outputting a single multi-scale feature map) can improve learning without increasing inference latency.

Summary by alphaXiv.