Reproduce “FS-I2P:A Hierarchical Focus-Sweep Registration Network with Dynamically Allocated Depth” 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

Zhixin Cheng, Yujia Chen, Xujing Tao, Bohao Liao, Xiaotian Yin, Baoqun Yin, Tianzhu Zhang

A hierarchical image-to-point cloud registration network named FS-I2P was developed, employing a Mamba-based "Focus–Sweep" interaction paradigm and a reinforcement learning-driven dynamic layer allocation strategy. The network achieved a 72.9% registration recall on RGB-D Scenes V2, outperforming previous methods by 4.5 to 16.5 percentage points, and demonstrated enhanced robustness and efficiency across indoor and outdoor benchmarks.

Key insights

  • A cognitively inspired "Focus–Sweep" interaction pattern, implemented with Mamba (State Space Models), offers a stable and efficient alternative to transformers for cross-modal feature aggregation in I2P registration.
  • Dynamically adjusting the depth of interaction layers via reinforcement learning significantly enhances robustness by tailoring computational refinement to the specific scene complexity and matching challenges.
  • Integrating global context modeling (Focus) with local structural propagation (Sweep) effectively manages scale ambiguity and refines correspondences across diverse feature resolutions.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.