Reproduce “LiDAR-based Panoptic Segmentation via Dynamic Shifting Network” 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
Fangzhou Hong, Hui Zhou, Xinge Zhu, Hongsheng Li, Ziwei Liu
DS-Net proposes a framework for LiDAR-based panoptic segmentation that unifies semantic and instance predictions for autonomous driving scenes. The system achieved the first position on the SemanticKITTI public leaderboard, surpassing the second-place method by 2.6% in Panoptic Quality (PQ), and outperformed baselines on nuScenes by 2.4% PQ.
Key insights
- The learnable dynamic shifting approach, which adapts kernel functions on-the-fly, is crucial for effectively clustering noisy and non-uniform LiDAR point distributions, outperforming fixed-parameter heuristic methods.
- Weighting across multiple pre-defined bandwidth candidates for the kernel function is more robust and effective than attempting to directly regress a single bandwidth per point.
- The model demonstrates interpretability, as the average learned bandwidths from the dynamic shifting module are roughly proportional to the actual sizes of the instances.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.