Reproduce “LiSu: A Dataset and Method for LiDAR Surface Normal Estimation” on OpenResearch

About this paper

Dušan Malić, Christian Fruhwirth-Reisinger, Samuel Schulter, Horst Possegger

Researchers from Graz University of Technology and Amazon introduce LiSu, the first large-scale synthetic LiDAR dataset featuring ground truth surface normal annotations, alongside a novel deep learning method. This method achieves state-of-the-art performance with a mean angular error of 6.30 degrees and processes LiDAR frames in approximately 90 milliseconds, demonstrating robust synthetic-to-real domain adaptation and improving downstream neural surface reconstruction.

Key insights

• Large-scale synthetic datasets with high-fidelity ground truth can serve as a crucial foundation for training robust deep learning models for LiDAR tasks, alleviating the need for costly manual annotation.
• Explicitly enforcing spatial and temporal coherence, along with unit-norm constraints through regularization terms, significantly enhances the accuracy and robustness of surface normal estimation for sparse and noisy LiDAR point clouds.
• Efficient architectures like Point Transformer V3, which process entire point clouds in a single pass, can achieve substantial speedups (e.g., from >14 seconds to ~90 milliseconds per frame) compared to patch-based methods, critical for real-time applications.

Summary by alphaXiv.