Reproduce “LION: Linear Group RNN for 3D Object Detection in Point Clouds” 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
Zhe Liu, Jinghua Hou, Xinyu Wang, Xiaoqing Ye, Jingdong Wang, Hengshuang Zhao, Xiang Bai
The LION framework introduces a linear group RNN-based approach for 3D object detection in point clouds, enabling efficient long-range relationship modeling in much larger feature groups. This approach establishes new state-of-the-art performance on key autonomous driving benchmarks, including Waymo (e.g., 1.9 mAPH/L2 gain over DSVT-Voxel) and nuScenes (e.g., 1.2 NDS gain over DSVT).
Key insights
- Linear RNNs can serve as an effective and efficient backbone for 3D point cloud perception, challenging the dominance of transformers in this domain.
- Enabling feature interaction in much larger groups (thousands of voxels) through the linear complexity of RNNs is crucial for scalable long-range modeling in sparse 3D point clouds.
- Explicitly integrating local spatial context, such as with a 3D Spatial Feature Descriptor, is vital when linearizing 3D data for sequential models like RNNs to mitigate spatial information loss.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.