Reproduce “iSDF: Real-Time Neural Signed Distance Fields for Robot Perception” 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
Joseph Ortiz, Alexander Clegg, Jing Dong, Edgar Sucar, David Novotny, Michael Zollhoefer, Mustafa Mukadam
Researchers at Imperial College London and Meta AI developed iSDF, a real-time continual learning framework that reconstructs non-truncated Signed Distance Fields (SDFs) using neural networks from depth images for robot perception. The system produces more accurate, complete, and coherent SDFs than traditional voxel-based methods, while maintaining a low memory footprint of approximately 1MB and achieving an average iteration time of 33ms.
Key insights
- Neural fields can effectively learn and represent complete, non-truncated SDFs in real-time across room-scale environments, surpassing traditional voxel-grid methods in terms of map completeness, memory efficiency, and adaptive detail.
- A carefully engineered self-supervision strategy, particularly using batch distance to approximate closest surface points, is critical for accurately learning signed distance values throughout the 3D space, including unobserved regions.
- The inherent properties of continuous neural representation, combined with an active sampling scheme and a tailored loss function (including a modified Eikonal regularization), enable robust denoising, interpolation, and dynamic adaptation of detail.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.