Reproduce “DXQ-Net: Differentiable LiDAR-Camera Extrinsic Calibration Using Quality-aware Flow” 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
Xin Jing, Xiaqing Ding, Rong Xiong, Huanjun Deng, Yue Wang
DXQ-Net presents an end-to-end differentiable framework for LiDAR-camera extrinsic calibration that estimates both calibration flow and its uncertainty. The method achieves state-of-the-art rotational accuracy (0.084 degrees) and competitive translational accuracy (1.425 cm) on KITTI datasets while demonstrating strong generalization to different cameras and environments due to its quality-aware probabilistic modeling.
Key insights
- Integrating a differentiable geometric optimization module provides a crucial inductive bias, significantly improving generalization and accuracy for deep learning-based calibration.
- Modeling calibration flow probabilistically allows the network to learn and leverage uncertainty, enhancing robustness by identifying and down-weighting unreliable data associations.
- End-to-end differentiability through the entire pipeline, including pose estimation, enables direct optimization for the final extrinsic error, leading to better overall performance.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.