Research map / 3D vision / Gaussian splatting
ClusterSplat: Semantic Cluster Selection for 3D Visual Grounding in Gaussian Splatting
Jaesung Lee, Hyeontaek Hwang, Gabriel Manalu, Daeyoung Kim
Read the paperCheck your idea against 137,217 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 ClusterSplat, which selects semantic clusters of 3D Gaussians for referring-expression grounding in 3DGS instead of heuristic thresholding of primitive-level relevance scores.
Problem
3D visual grounding in 3D Gaussian Splatting must output both a 2D mask and a set of 3D Gaussians for a referring expression; prior methods use heuristic thresholds or top-ratio filtering on primitive relevance scores, which vary across scenes and prompts and make such heuristics unreliable.
Approach
Replaces per-primitive thresholding with selection over semantic clusters of Gaussians, so the target set is chosen at the cluster level rather than by a scene-specific cutoff on dense relevance scores.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace primitive-level top-ratio filtering with scene-adaptive cluster-level Top-1 selection | ClusterSplat vs ReferSplat · ScanRefer 3D | mIoU: 4.05 -> 23.33 | Improved |
| train ReferSplat with our loss (same supervision) | ReferSplat · ScanRefer 2D | mIoU: 31.30 -> 31.55 | No clear effect |
| use scene-adaptive clustering instead of fixed KMeans K=128 (ablation) | ClusterSplat cluster candidates · ScanRefer 3D mIoU | mIoU: 18.31 -> 23.33 | Improved |
| increase fixed KMeans K from 128 to 512 (ablation) | fixed KMeans clusters · ScanRefer 3D mIoU | mIoU: 18.31 -> 10.97 | Hurt |
| use description-length merge instead of fixed KMeans + pair merge (ablation) | cluster construction · ScanRefer 3D mIoU | mIoU: 21.23 -> 23.33 | Improved |
| add xyz cost to the description-length merge (ablation) | scene-adaptive clustering · ScanRefer 3D mIoU | mIoU: 23.33 -> 21.63 | Hurt |
| select Top-5 cluster union instead of Top-1 (ablation) | scene-adaptive clustering · ScanRefer 3D mIoU | mIoU: 23.33 -> 14.96 | Hurt |
| use multi-cluster Top-k union with fixed KMeans K=512 (ablation) | fixed KMeans K=512 · ScanRefer 3D mIoU | mIoU: 10.97 -> 14.62 | Improved |
| use frozen CLIP instead of frozen BERT as text encoder (ablation) | ClusterSplat scorer · ScanRefer 3D mIoU | mIoU: 14.62 -> 23.33 | Improved |
| use local VLM caption generators instead of GPT-4.1 (ablation) | pseudo-caption source · ScanRefer 3D mIoU | mIoU: 23.33 -> 21.23 (InternVL3.5-8B), 20.01 (Qwen3-VL-8B) | Hurt |
| use cluster-level selection instead of primitive filtering on class-name queries | ClusterSplat vs OpenGaussian · ScanNet v2 class name 3D | mIoU: 31.05 -> 29.21 mIoU; mAcc 46.39 -> 46.60 | Mixed |
| use cluster-level selection on Ref-LERF versus ReferSplat with our loss | ClusterSplat · Ref-LERF average | mIoU: 26.59 -> 31.14 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- ReferSplat — ReferSplat: Referring Segmentation in 3D Gaussian Splatting (ICML 2025) · better on 2 of 2 shared results
- LangSplat — LangSplat: 3D Language Gaussian Splatting (CVPR 2024) · better on 2 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- SceneSplat++: A Large Dataset and Comprehensive Benchmark for Language Gaussian Splatting (NeurIPS 2025 · similarity 0.91)
- Bootstraping Clustering of Gaussians for View-consistent 3D Scene Understanding (AAAI 2025 · similarity 0.90)
- 3D Student Splatting and Scooping (CVPR 2025 · similarity 0.89)
- GaussianSpa: An "Optimizing-Sparsifying" Simplification Framework for Compact and High-Quality 3D Gaussian Splatting (CVPR 2025 · similarity 0.89)
- Efficient Decoupled Feature 3D Gaussian Splatting via Hierarchical Compression (CVPR 2025 · similarity 0.89)
- AdpSplit: Error-Driven Adaptive Splitting for Faster Geometry Discovery in 3D Gaussian Splatting (NeurIPS 2026 · similarity 0.89)
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.
