GaussianSelector: Lightweight Human-Guided Object Selection in 3D Gaussian Splatting with Graph Optimization

Analysis

Why This Paper Matters

GaussianSelector addresses a critical bottleneck in 3D scene editing: selecting a complete object from a reconstructed scene with minimal user effort. Existing methods either require retraining the Gaussian representation to embed per-object labels, which is computationally heavy, or build dense multi-view SAM observations that demand dense viewpoint coverage rarely available in practice. This paper introduces a training-free framework that works directly on native Gaussian primitives, making it highly practical for real-world scenarios where computational resources and viewpoint coverage are limited.

The significance lies in its ability to achieve competitive selection quality with sparse inputs—both sparse views and sparse scribble guidance. This is a major step forward for interactive 3D editing, as it reduces the user burden and computational cost, enabling more accessible and efficient workflows. The method's design also supports multi-round refinement, allowing users to iteratively correct selections, which is essential for practical deployment.

Technical Contributions

• Superpoint Coarsening: Dense Gaussians are coarsened into geometrically coherent superpoints, reducing complexity while preserving structural integrity.
• Continuity-Weighted Graph: A graph is constructed using appearance and spatial cues, with weights reflecting continuity, enabling effective propagation of sparse evidence.
• Visibility-Aware Transmittance Coverage: Sparse user scribbles are lifted into 3D using visibility-aware transmittance, ensuring accurate correspondence between 2D scribbles and 3D primitives.
• Global Graph-Cut Energy Minimization: Selection is formulated as a global energy minimization problem, solved via graph-cut, which propagates sparse evidence to a complete 3D object.
• Multi-Round Refinement: The framework naturally supports iterative user corrections from additional viewpoints, progressively improving selection quality.

Results

While the abstract does not provide specific numerical metrics, it states that GaussianSelector achieves competitive selection quality against state-of-the-art multi-view SAM-based methods. The key advantages are significantly fewer interaction views and substantially lower computational overhead. This suggests that the method is not only effective but also efficient, making it suitable for real-time or near-real-time interactive applications.

Significance

The broader impact of GaussianSelector lies in its potential to democratize 3D scene editing and asset extraction. By eliminating the need for retraining and dense multi-view coverage, it lowers the barrier to entry for practitioners who may not have access to high-end computational resources. This could accelerate adoption in fields like robotics, augmented reality, and virtual production, where quick and accurate object selection is crucial. Furthermore, the training-free nature of the method means it can be applied to any existing 3DGS reconstruction without additional training, making it a versatile tool for the AI community.