Research map / 3D vision / Gaussian splatting
SceneSplat++: A Large Dataset and Comprehensive Benchmark for Language 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.
Introduces SceneSplat-Bench, a large-scale 3D benchmark for language Gaussian Splatting methods across 1060 scenes, and SceneSplat-49K, a ~49K-scene 3DGS dataset, showing the generalizable feed-forward SceneSplat beats per-scene methods.
Problem
Existing language Gaussian Splatting (LGS) methods are evaluated on few scenes, close to training views, and mostly in 2D, limiting conclusions about holistic 3D scene understanding; large-scale 3DGS scene datasets are also scarce.
Approach
A 3D-native benchmark that evaluates per-scene optimization-based, per-scene optimization-free, and generalizable LGS methods on many scenes with foreground mIoU/mAcc on 3D Gaussians, paired with a large curated 3DGS scene dataset (SceneSplat-49K) used to scale a generalizable model.
Key results
- Generalizable SceneSplat reaches 0.3540 f-mIoU on ScanNet20 (312 scenes), outperforming the best per-scene method GOI at 0.2115.
- On the ScanNet++ 10-scene subset, SceneSplat achieves 0.2263 f-mIoU vs. 0.1593 for GOI, while the best 3D localization accuracy on ScanNet is 0.7701 for SceneSplat vs. 0.6750 for LUDVIG.
- Scaling indoor training data from 280 to 3503 scenes raises ScanNet++ f-mIoU from 0.168 to 0.284.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Train on 3503 indoor scenes instead of 280 ScanNet++ scenes | SceneSplat · ScanNet++ | f-mIoU: 0.168 → 0.284 | Helped |
| Train on 3503 indoor scenes instead of 280 ScanNet++ scenes | SceneSplat · ScanNet200 | f-mIoU: 0.108 → 0.165 | Helped |
| Use larger indoor training data (3503 scenes) instead of the 280-scene baseline | SceneSplat · HoliCity | mIoU: 0.263 mIoU; 7.4% over the 280-scene baseline | Helped |
| Train with HoliCity-specific data instead of indoor-only data | SceneSplat · HoliCity | mIoU: 0.2633 → 0.2880 | Helped |
| Use SceneSplat rather than Gradient-Weighted 3DGS | 3D zero-shot segmentation · ScanNet++ (10 scenes) | f-mIoU: 0.1593 → 0.2263 | Helped |
| Use SceneSplat rather than Gradient-Weighted 3DGS | 3D zero-shot segmentation · Matterport3D (10 scenes) | f-mIoU: 0.3083 → 0.2732 | Hurt |
| Use SceneSplat rather than Gradient-Weighted 3DGS | 3D zero-shot segmentation · ScanNet20 (10 scenes) | f-mIoU: 0.4184 → 0.2968 | Hurt |
| Use SceneSplat rather than OccamLGS | 3D zero-shot segmentation · ScanNet200 (10 scenes) | f-mIoU: 0.1936 → 0.1387 | Hurt |
| Use SceneSplat rather than per-scene baseline methods | 3D zero-shot segmentation · HoliCity (10 scenes) | f-mIoU: 0.2479 → 0.3081 | Helped |
| Use SceneSplat rather than Gradient-Weighted 3DGS | 3D zero-shot segmentation · ScanNet20 (312 scenes) | f-mIoU: 0.3573 → 0.3540 | No clear effect |
Surprising findings
- An indoor-only trained SceneSplat transfers zero-shot to outdoor HoliCity scenes, reaching 0.263 mIoU without outdoor training data.
- Per-scene optimization-free methods generally outperform per-scene optimization-based methods in 3D segmentation and runtime, and the authors conclude per-scene optimization is unnecessary.
Limitations
- Outdoor benchmark is limited in number of semantic classes (HoliCity has 4 classes).
- Not all 3DGS scenes in the dataset include precomputed vision-language embeddings (about 12K of 49K).
- Dataset scale, while large, could be further expanded.
Experimental setup
Datasets: SceneSplat-7K DL3DV-10K HoliCity Aria Synthetic Environments crowdsourced 3DGS scenes ScanNet ScanNet++ Matterport3D ScanNet200 LERF 3DOVS
Benchmarks: SceneSplat-Bench (ScanNet20, ScanNet200, ScanNet++, Matterport3D, HoliCity) LERF-style 2D evaluation 3D object localization on ScanNet and ScanNet++
Metrics: f-mIoU f-mAcc mIoU mAcc localization accuracy runtime
Baselines: LangSplat OpenGaussian Feature-3DGS FMGS GOI Gradient-Weighted 3DGS LUDVIG OccamLGS SceneSplat
Builds on: 3D Gaussian Splatting LangSplat SceneSplat gsplat Markov Chain Monte Carlo densification
Closest papers by meaning
- PlanarGS: High-Fidelity Indoor 3D Gaussian Splatting Guided by Vision-Language Planar Priors (NeurIPS 2025 · similarity 0.92)
- Dr. Splat: Directly Referring 3D Gaussian Splatting via Direct Language Embedding Registration (CVPR 2025 · similarity 0.91)
- ReferSplat: Referring Segmentation in 3D Gaussian Splatting (ICML 2025 · similarity 0.91)
- LangSplat: 3D Language Gaussian Splatting (CVPR 2024 · similarity 0.91)
- Segment then Splat: Unified 3D Open-Vocabulary Segmentation via Gaussian Splatting (NeurIPS 2025 · similarity 0.91)
- Efficient Decoupled Feature 3D Gaussian Splatting via Hierarchical Compression (CVPR 2025 · similarity 0.91)
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.