Research map / 3D vision / 3D generation
InternScenes: A Large-scale Simulatable Indoor Scene Dataset with Realistic Layouts
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 InternScenes, a ~40K-scene simulatable indoor dataset merging real-scan, procedurally generated, and designer-created scenes with dense small objects and physics-based collision resolution, and benchmarks layout generation and point-goal navigation on it.
Problem
Existing 3D indoor scene datasets lack scale, diversity, realistic layouts with small items, and simulation-readiness (collisions, non-interactive objects, non-simulatable scans).
Approach
Combine three scene sources (EmbodiedScan real scans converted to sim replicas via asset retrieval, Infinigen Indoors procedural scenes, and designer-created scenes annotated by hand), keep many small objects, substitute ~20% of assets with PartNet-Mobility interactive objects, and resolve collisions via bounding-box optimization plus SAPIEN physics simulation.
Key results
- InternScenes contains about 40K scenes, 48K regions across 15 types, 1.96M objects from 288 classes, and an average of 41.5 objects per region.
- On point-goal navigation, NavDP fine-tuned on InternScenes data reaches 51.0% success on Real2Sim vs 48.3% for NavDP and 23.6% for DD-PPO.
- Layout generation baselines (ATISS, DiffuScene, PhyScene) show worse FID on the full version than on the simplified version; e.g., PhyScene resting FID 88.02 full vs 23.78 simplified... (per-method comparisons reported in the table).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Train on the full version with small objects rather than the simplified version (ablation) | ATISS, DiffuScene, and PhyScene · InternScenes scene generation, Resting, Living, and Dining regions | FID: 22.88 -> 88.02; 23.54 -> 66.59; 26.76 -> 122.95 | Hurt |
| Train on the full version with small objects rather than the simplified version (ablation) | ATISS, DiffuScene, and PhyScene · InternScenes scene generation, Resting, Living, and Dining regions | CKL | Mixed |
| Fine-tune NavDP on 118,784 InternScenes trajectories (ablation) | NavDP · InternScenes-Real2Sim | Success: 48.3 -> 51.0 | Helped |
| Fine-tune NavDP on 118,784 InternScenes trajectories (ablation) | NavDP · InternScenes-Real2Sim | SPL: 45.3 -> 49.4 | Helped |
| Fine-tune NavDP on 118,784 InternScenes trajectories (ablation) | NavDP · InternScenes-Gen | Success: 61.9 -> 63.6 | Helped |
| Fine-tune NavDP on 118,784 InternScenes trajectories (ablation) | NavDP · InternScenes-Gen | SPL: 61.8 -> 61.7 | No clear effect |
Surprising findings
- Baseline layout generators degrade sharply when small objects are included, with FID rising from roughly 23-31 on the simplified set to over 100 on the full set.
- Fine-tuning NavDP on 118K InternScenes trajectories yields only a slight gain, suggesting model capacity is a bottleneck.
Limitations
- The pipeline still relies on manual annotation for region and instance labeling.
- Scene diversity can be further improved according to the authors.
- The navigation benchmark uses only 30 scenes and 20 episodes per scene, a small evaluation set.
Experimental setup
Datasets: EmbodiedScan Infinigen Indoors generated scenes designer-created scenes Objaverse PartNet-Mobility Cap3D
Benchmarks: InternScenes layout generation (full and simplified versions) InternScenes-Real2Sim point-goal navigation InternScenes-Gen point-goal navigation
Metrics: FID KID SCA CKL Success rate SPL FPS CPU/GPU usage
Baselines: ATISS DiffuScene PhyScene DD-PPO NavDP
Builds on: EmbodiedScan Infinigen Indoors Objaverse PartNet-Mobility SAPIEN
Closest papers by meaning
- Direct Numerical Layout Generation for 3D Indoor Scene Synthesis via Spatial Reasoning (NeurIPS 2025 · similarity 0.88)
- SceneWeaver: All-in-One 3D Scene Synthesis with an Extensible and Self-Reflective Agent (NeurIPS 2025 · similarity 0.88)
- Habitat Synthetic Scenes Dataset (HSSD-200): An Analysis of 3D Scene Scale and Realism Tradeoffs for ObjectGoal Navigation (CVPR 2024 · similarity 0.88)
- InstructScene: Instruction-Driven 3D Indoor Scene Synthesis with Semantic Graph Prior (ICLR 2024 · similarity 0.88)
- Image2Sim: Scaling Embodied Navigation via Generative Neural Simulator (NeurIPS 2026 · similarity 0.88)
- S-INF: Towards Realistic Indoor Scene Synthesis via Scene Implicit Neural Field (AAAI 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.