GaussianSelector: Lightweight Human-Guided Object Selection in 3D Gaussian Splatting with Graph Optimization
GaussianSelector turns sparse user scribbles into complete 3D object selections without retraining the Gaussian scene.
The method works directly on native 3D Gaussian primitives, grouping them into coherent superpoints and solving selection with a graph-cut optimization. User scribbles are projected into 3D with visibility-aware coverage, then propagated across the graph. The authors say it supports iterative correction from additional viewpoints. In experiments, it matched multi-view SAM-based selection quality while using fewer interaction views and less computation. HF Daily Papers' note
The method works directly on native 3D Gaussian primitives, grouping them into coherent superpoints and solving selection with a graph-cut optimization. User scribbles are projected into 3D with visibility-aware coverage, then propagated across the graph. The authors say it supports iterative correction from additional viewpoints. In experiments, it matched multi-view SAM-based selection quality while using fewer interaction views and less computation. HF Daily Papers' note
score 4