Reproduce “Efficient Spatial-Temporal Information Fusion for LiDAR-Based 3D Moving Object Segmentation” 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
Jiadai Sun, Yuchao Dai, Xianjing Zhang, Jintao Xu, Rui Ai, Weihao Gu, Xieyuanli Chen
A deep learning method for online LiDAR-based 3D Moving Object Segmentation integrates a dual-branch network for spatio-temporal feature fusion with a coarse-to-fine 3D refinement module. The approach achieved a 70.16% IoU on the SemanticKITTI-MOS hidden test set, outperforming prior methods while operating at real-time speeds of 85.25 ms per scan.
Key insights
- Dedicated branches for spatial appearance and temporal motion, combined with attention mechanisms, enable more effective spatio-temporal information fusion than simple concatenation.
- Leveraging both 2D range image efficiency and 3D point cloud geometry through a coarse-to-fine refinement module significantly mitigates boundary-blurring artifacts.
- Augmenting training data with manually refined labels from diverse environments improves the generalization and robustness of MOS models.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.