Reproduce “GSMap: 2D Gaussians for Online HD Mapping” 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
Zhenxuan Zeng, Lingxuan Wang, Sheng Yang, Yanan He, Mingxia Chen, Wei Suo, Peng Wang
GSMap introduces a framework for online HD map construction that uses ordered sequences of learnable 2D Gaussian primitives to represent map elements, unifying rasterized geometry and vectorized topology. This approach achieved improved Chamfer distance-based and IoU-based accuracy on nuScenes and Argoverse2 datasets, outperforming prior hybrid methods while maintaining inference efficiency.
Key insights
- Learnable 2D Gaussian primitives can serve as a unified, differentiable representation for HD map elements, intrinsically bridging the gap between dense geometric (rasterized) and structured topological (vectorized) information.
- Simultaneous supervision from both differentiable rasterization (for dense pixel-level feedback) and direct vectorization (for inherent structural consistency) on a shared Gaussian representation enhances map accuracy and robustness.
- The proposed Gaussian-based framework is generalizable, improving performance when integrated with different existing BEV perception architectures like MapTR and MapTRv2 without significant overhead.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.