Reproduce “LIE: LiDAR-only HD Map Construction with Intensity Enhancement via Online Knowledge Distillation” 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
Kanak Mazumder, Fabian B. Flohr
LIE (LiDAR-only HD Map Construction with Intensity Enhancement) provides a real-time framework for constructing high-definition maps by employing an online Intensity-to-LiDAR knowledge distillation strategy. The framework achieves 58.0% mIoU on nuScenes, outperforming all single-modality camera-based methods and several fusion approaches, while operating at over 36 FPS.
Key insights
- Underutilized LiDAR intensity information provides rich structural and reflectivity cues, which can be effectively leveraged for dense semantic segmentation in HD maps.
- An online "Intensity-to-LiDAR" knowledge distillation strategy successfully transfers dense semantic knowledge from a specialized intensity-map teacher to a sparse LiDAR-only student, enabling enhanced feature learning without additional runtime sensor inputs.
- LiDAR-only systems, when augmented with this intensity-driven knowledge, can achieve real-time HD map construction performance that is competitive with or exceeds camera-based and many fusion-based methods, offering superior robustness and generalization.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.