Reproduce “Scene Recomposition by Learning-based ICP” 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
Hamid Izadinia, Steven M. Seitz
A system for 3D scene recomposition automatically constructs environment models from high-quality CAD shapes using RGBD input. Its core is Learning-based ICP (LICP), a reinforcement learning approach that robustly aligns CAD models to noisy, occluded 3D scans, outperforming prior methods in 6DoF pose estimation on real and synthetic data while training solely on synthetic scenes.
Key insights
- Deep reinforcement learning, guided by a non-differentiable ICP-based reward, can effectively learn robust 6DoF object alignment from partial and cluttered scans.
- Training a 3D alignment model solely on synthetic data from virtual environments can successfully generalize to real-world scenes, significantly reducing data annotation requirements.
- Scene recomposition offers an alternative to traditional geometric reconstruction, providing complete, cleaner, and semantically richer 3D representations by leveraging high-quality CAD models.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.