Reproduce “GaussianCut: Interactive segmentation via graph cut for 3D Gaussian Splatting” on OpenResearch

About this paper

Umangi Jain, Ashkan Mirzaei, Igor Gilitschenski

GaussianCut introduces an interactive method for segmenting objects within existing 3D Gaussian Splatting (3DGS) models, leveraging a graph cut algorithm to convert noisy 2D user inputs into precise, 3D-consistent object masks without requiring any additional training or modification of the 3DGS scene. The approach demonstrated superior performance on benchmarks like the NVOS dataset, achieving a 92.5% IoU and improving object rendering PSNR by 4.05 dB compared to prior methods.

Key insights

• Graph cut optimization effectively refines noisy 2D-derived masks into coherent and precise 3D segmentations for 3DGS models by enforcing 3D consistency.
• Integrating both spatial and color similarities of Gaussians, alongside robust initial likelihoods, within a graph energy function is crucial for achieving high-quality 3D segmentation.
• Operating on *pretrained* 3DGS models enables a flexible, training-free approach that avoids the computational overhead and additional training time of feature-based segmentation methods.

Summary by alphaXiv.