Research map / 3D vision / Gaussian splatting
GaussianCut: Interactive segmentation via graph cut for 3D Gaussian Splatting
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes GaussianCut, a training-free interactive segmentation method that propagates 2D video-segmentation masks to 3D Gaussians and refines them with a graph cut over a Gaussian k-NN graph.
Problem
Interactive selection and segmentation of objects in scenes represented by 3D Gaussian Splatting, without modifying the 3DGS optimization or requiring feature-learning or segmentation-aware training.
Approach
Builds a graph whose nodes are the 3D Gaussians, with spatial and color n-link weights and t-link weights derived from multi-view mask-based opacity-transmittance ratios plus cluster similarity to high-confidence foreground/background Gaussians, then partitions it with min-cut on an optimized pretrained 3DGS model.
Key results
- On the NVOS LLFF scenes, GaussianCut reaches 92.5% IoU and 98.4% accuracy, above SAGA (90.9% IoU) and SA3D (90.3% IoU).
- On SPIn-NeRF, GaussianCut achieves 92.9% mean IoU versus 91.9% for SA3D and 90.4% for MVSeg.
- On 3D-OVS, GaussianCut averages 96.8% on bed scenes but is outperformed by Gaussian Grouping on lawn (89.8% vs 96.5%).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add graph cut to coarse splatting (ablation) | GaussianCut · NVOS | IoU: 91.2 -> 92.5 | Helped |
| remove n-link spatial similarity (ablation) | GaussianCut · NVOS | IoU: 92.5 -> 92.4 | Hurt |
| remove n-link color similarity (ablation) | GaussianCut · NVOS | IoU: 92.5 -> 92.3 | Hurt |
| remove t-link cluster similarity (ablation) | GaussianCut · NVOS | IoU: 92.5 -> 91.5 | Hurt |
| increase masks from 5 to 43 (ablation) | GaussianCut on Fortress · Fortress | IoU: 97.7 -> 97.9 | Helped |
| increase masks from 11 to 102 (ablation) | GaussianCut on Lego · Lego | IoU: 87.3 -> 89.2 | Helped |
| increase neighbors from 1 to 100 (ablation) | GaussianCut on Horns · Horns | IoU: 91.9 -> 94.3 | Helped |
| replace soft assignment with hard binary weights (ablation) | GaussianCut on Fern · Fern (NVOS) | IoU: 83.06 -> 82.56 | Hurt |
| use user scribbles directly with graph cut rather than coarse splatting from scribbles (ablation) | GaussianCut on Fern · Fern | IoU: 8.17 -> 47.97 | Helped |
| use GaussianCut instead of the strongest reported segmentation baseline | GaussianCut · NVOS | IoU (%): 90.9 -> 92.5 | Helped |
| use GaussianCut instead of SA3D | GaussianCut · SPIn-NeRF | IoU (%): 92.4 -> 92.9 | Helped |
| use GaussianCut instead of Coarse Splatting | GaussianCut · Shiny | IoU (%): 94.3 -> 95.0 | Helped |
Surprising findings
- Feature-based 3DGS methods such as Gaussian Grouping and LangSplat outperform GaussianCut on some scenes, which the authors attribute to limited interactivity when optimized features do not delineate the object.
- Graph cut is slow: segmentation takes about 88.8 s on average on NVOS, much longer than the feature-based baselines' segmentation times.
Limitations
- Segmentation is not real-time; graph cut can take up to a few minutes.
- Performance depends on the quality of the 2D image/video segmentation models, and partial recovery occurs when the initial mask is far off.
- Hyperparameters (lambda, gamma, thresholds) are scene-dependent and tuned per dataset.
Experimental setup
Datasets: LLFF Shiny SPIn-NeRF 3D-OVS Mip-NeRF 360 Tanks and Temples LERF
Benchmarks: NVOS SPIn-NeRF Shiny 3D-OVS
Metrics: IoU Accuracy PSNR SSIM LPIPS
Baselines: NVOS ISRF SA3D SAGA SAGD Gaussian Grouping LangSplat MVSeg Contrastive Gaussian Clustering
Builds on: 3D Gaussian Splatting Boykov-Kolmogorov min-cut SAM-Track Boykov and Jolly interactive graph cut
Closest papers by meaning
- SceneSplat++: A Large Dataset and Comprehensive Benchmark for Language Gaussian Splatting (NeurIPS 2025 · similarity 0.88)
- ReferSplat: Referring Segmentation in 3D Gaussian Splatting (ICML 2025 · similarity 0.88)
- Gaussian Graph Network: Learning Efficient and Generalizable Gaussian Representations from Multi-view Images (NeurIPS 2024 · similarity 0.88)
- iSegMan: Interactive Segment-and-Manipulate 3D Gaussians (CVPR 2025 · similarity 0.88)
- Rethinking End-to-End 2D to 3D Scene Segmentation in Gaussian Splatting (CVPR 2025 · similarity 0.88)
- Segment then Splat: Unified 3D Open-Vocabulary Segmentation via Gaussian Splatting (NeurIPS 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.