Reproduce “Relation-Centric Open-Vocabulary 3D Gaussian Segmentation” 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
Eunsung Cha, Hyunjoon Lee, Jaesik Park
PairGS groups 3D scene elements into queryable objects by comparing pairwise evidence across views, using SAM-masked CLIP features to propose candidate neighbors and a hierarchy to organize clusters. On LERF-OVS, it reports mean mIoU of 60.4 across four scenes, versus 52.2 for LaGa. It cannot recover objects that no view’s masks separate.
Key insights
- The 3D Gaussian representation inherently provides sufficient cues, such as view-dependent contribution weights and multi-view mask evidence, to estimate robust pairwise relations between Gaussians.
- Explicitly modeling these pairwise relations allows for more coherent and fine-grained instance separation compared to relying solely on individual Gaussian features or extensive per-scene optimization.
- A hierarchical clustering approach, built upon localized affinity relationships, can effectively support multi-granular open-vocabulary queries while maintaining computational efficiency.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.