Conceitos Básicos
Proposing the Voxel-Cross-Pixel (VXP) approach for accurate image-LiDAR place recognition, surpassing state-of-the-art methods.
Resumo
The study introduces the VXP approach to address challenges in global place recognition using images and LiDAR data. It consists of a two-stage training process focusing on local and global descriptors. Extensive experiments demonstrate superior performance on various datasets compared to existing methods.
Directory:
- Introduction
- Challenges in global place recognition due to GPS signal outages.
- Importance of onboard devices like cameras and LiDARs for autonomous driving.
- Related Work
- Overview of uni-modal and fused-modal place recognition techniques.
- Method
- Description of the VXP pipeline for image-LiDAR place recognition.
- Experiments and Results
- Evaluation on Oxford RobotCar, ViViD++, and KITTI Odometry datasets.
- Ablation Studies
- Comparison of one-stage vs. two-stage descriptor optimization.
- Conclusion
- Summary of the proposed VXP approach's effectiveness in image-LiDAR place recognition.
Estatísticas
"Extensive experiments on the three benchmarks (Oxford RobotCar, ViViD++ and KITTI) demonstrate our method surpasses the state-of-the-art cross-modal retrieval by a large margin."
"Our model achieves double accuracy on Top@1 for both 2D-3D and 3D-2D retrieval tasks compared to baseline methods."