LiSA: LiDAR Localization with Semantic Awareness

Abstract

LiDAR localization is a fundamental task in robotics and computer vision, estimating the pose of a LiDAR point cloud within a global map. Scene Coordinate Regression (SCR) has demonstrated state-of-the-art performance in this task, representing a scene as a neural network that outputs world coordinates for each input point. However, SCR treats all points equally, ignoring that not all objects are beneficial for localization — dynamic objects and repeating structures often negatively impact accuracy. We introduce LiSA, the first method that incorporates semantic awareness into SCR to boost localization robustness and accuracy. To avoid extra computation during inference, we distill knowledge from a segmentation model into the SCR network. Experiments show superior performance of LiSA on standard LiDAR localization benchmarks. Applying knowledge distillation not only preserves efficiency but also achieves higher accuracy than introducing explicit semantic segmentation modules.

Resources

arXiv: 2406.01843

Citation

@inproceedings{yang2024lisa,
  title     = {{LiSA}: {LiDAR} Localization with Semantic Awareness},
  author    = {Yang, Bochun and Li, Zijun and Li, Wen and Cai, Zhipeng and Wen, Chenglu and Zang, Yu and M{\"{u}}ller, Matthias and Wang, Cheng},
  booktitle = {IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  note      = {Highlight},
  year      = {2024}
}