Reproduce “Flow4D: Leveraging 4D Voxel Network for LiDAR Scene Flow Estimation” 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
Jaeyeul Kim, Jungwan Woo, Ukcheol Shin, Jean Oh, Sunghoon Im
Flow4D presents a 4D voxel network for LiDAR scene flow estimation, explicitly extracting spatio-temporal features from multiple point clouds while employing Spatio-Temporal Decomposition Blocks (STDBs) for computational efficiency. The framework achieved a 45.9% reduction in mean Dynamic Normalized EPE on the Argoverse 2 test set, outperforming prior methods and maintaining real-time operation at 15.1 FPS.
Key insights
- Explicit spatio-temporal feature extraction in a 4D voxel space can substantially improve LiDAR scene flow estimation accuracy compared to implicit or 2D-based methods.
- Spatio-Temporal Decomposition Blocks (STDBs) effectively reduce the computational burden of 4D convolutions, making 4D feature processing feasible for real-time applications.
- Expanding the temporal context by incorporating more input frames (e.g., five) provides richer cues, leading to more accurate dynamic motion prediction.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.