Reproduce “Multi-View Pose-Agnostic Change Localization with Zero Labels” 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
Chamuditha Jayanga Galappaththige, Jason Lai, Lloyd Windrim, Donald Dansereau, Niko Suenderhauf, Dimity Miller
A multi-view, label-free, and pose-agnostic change localization method was developed that embeds change information into a 3D Gaussian Splatting representation. This approach achieves a 1.7x mIoU improvement over state-of-the-art baselines on multi-object scenes, demonstrates robust change detection under varying lighting, and accurately generates change masks for entirely unseen viewpoints.
Key insights
- Embedding change magnitude as a view-independent parameter (degree 0 spherical harmonics) in 3D Gaussians effectively suppresses view-dependent false positives like reflections and shadows.
- Initializing the change-aware 3DGS from the pre-change reference scene's Gaussians helps retain structural information, especially for removed objects, crucial for accurate change localization.
- The proposed Change-3DGS framework can act as a multi-view extension for existing 2D change mask generation methods, significantly boosting their performance by enforcing 3D consistency.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.