Reproduce “3DGS-CD: 3D Gaussian Splatting-based Change Detection for Physical Object Rearrangement” 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
Ziqi Lu, Jianbo Ye, John Leonard
3DGS-CD introduces the first 3D Gaussian Splatting-based method for detecting physical object rearrangements in 3D scenes from multi-view RGB images. It achieves up to 14% higher accuracy and is three orders of magnitude faster than prior NeRF-based techniques, performing robustly even with sparse post-change observations.
Key insights
- The rasterization-based rendering pipeline of 3D Gaussian Splatting provides a massive speed advantage over NeRF's ray-casting, enabling near real-time 3D change detection.
- Combining an efficient 3D scene representation (3DGS) with a zero-shot segmentation model (EfficientSAM) allows for robust, object-level change identification without prior object knowledge or extensive training data.
- A modular pipeline integrating classic computer vision techniques with modern radiance fields can achieve high accuracy and efficiency for complex 3D scene understanding tasks like object rearrangement, even from sparse input data.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.