Reproduce “End-to-End Vectorized HD-map Construction with Piecewise Bezier Curve” 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
Limeng Qiao, Wenjie Ding, Xi Qiu, Chi Zhang
Researchers at MEGVII Inc. introduced BeMapNet, an end-to-end, post-processing-free architecture for vectorized HD-map construction from multi-view camera inputs, utilizing unified piecewise B´ezier curves. This method improved upon existing state-of-the-art by 18.0 mAP on the NuScenes dataset, achieving 40.7 mAP with EfficientNet-B0 while also demonstrating higher inference speed and fewer parameters.
Key insights
- Piecewise B´ezier curves offer a compact and expressive parameterization for complex, 'shape-changeful' HD-map elements, enabling consistent-degree representation at the semantic level and dynamic pieces at the instance level.
- The IPM-PE Align Layer effectively integrates 3D geometric priors into Transformer position encoding, improving the accuracy of 2D-3D feature alignment and mitigating the limitations of traditional Inverse Perspective Mapping assumptions.
- A novel Point-Curve-Region (PCR) loss provides robust, multi-level supervision (point, curve, and region) for B´ezier curve prediction, demonstrating synergistic effects that significantly improve overall performance compared to individual loss components.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.