Research map / 3D vision / 3D generation
Habitat Synthetic Scenes Dataset (HSSD-200): An Analysis of 3D Scene Scale and Realism Tradeoffs for ObjectGoal Navigation
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 HSSD-200, a human-authored set of 211 synthetic 3D houses with 18,656 object models, and shows ObjectNav agents trained on 122 HSSD scenes generalize better zero-shot to real scans than agents trained on 10K ProcTHOR scenes.
Problem
Whether scale or realism of synthetic 3D scene datasets matters more for training ObjectGoal navigation agents that generalize zero-shot to real-world scanned environments.
Approach
A smaller but higher-fidelity, human-authored synthetic scene dataset with real-world object dimensions, semantic labels and realistic co-occurrence statistics transfers better to real scans than much larger procedurally generated datasets; scale benefits saturate quickly.
Key results
- Agents trained on HSSD-122 reach 19.15% zero-shot success on HM3DSem vs 12.53% for ProcTHOR-10K agents (SPL 7.71 vs 5.26).
- On MP3D, HSSD-122 agents achieve 12.56% success vs 8.26% for ProcTHOR-10K agents.
- HSSD has lower FID/KID to real images (e.g., FID 61.6 vs 87.2 ProcTHOR on MP3D) and higher co-occurrence similarity to HM3DSem (0.40 vs 0.14).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| train on HSSD-122 rather than ProcTHOR-10K | ObjectNav agent · HM3DSem val | success: 12.53 ± 0.49 -> 19.15 ± 0.95 | Helped |
| train on HSSD-122 rather than ProcTHOR-10K | ObjectNav agent · HM3DSem val | SPL: 5.26 ± 0.2 -> 7.71 ± 0.47 | Helped |
| train on HSSD-122 rather than ProcTHOR-10K | ObjectNav agent · MP3D val | success: 8.26 ± 0.73 -> 14.44 ± 1.29 | Helped |
| train on HSSD-122 rather than ProcTHOR-10K | ObjectNav agent · MP3D val | SPL: 2.96 ± 0.46 -> 5.17 ± 0.48 | Helped |
| use HSSD-122 rather than ProcTHOR-122 (ablation) | ObjectNav agent · HM3DSem val | success: 9.3 -> 19.2 | Helped |
| use HSSD-122 rather than ProcTHOR-122 (ablation) | ObjectNav agent · HM3DSem val | SPL: 3.5 -> 7.7 | Helped |
| use HSSD-122 rather than ProcTHOR-122 (ablation) | ObjectNav agent · MP3D val | success: 5.5 -> 14.4 | Helped |
| use HSSD-122 rather than ProcTHOR-122 (ablation) | ObjectNav agent · MP3D val | SPL: 1.8 -> 5.2 | Helped |
| increase ProcTHOR training scale from 122 scenes to 10K scenes (ablation) | ObjectNav agent · HM3DSem and MP3D val | zero-shot generalization | No clear effect |
Surprising findings
- Agents trained on 122 HSSD scenes outperform agents trained on 10,000 ProcTHOR scenes in zero-shot real-scan ObjectNav.
- Scaling ProcTHOR from 122 to 10K scenes yields only marginal ObjectNav improvement.
Limitations
- Only one agent type (monolithic pixels-to-actions RL) was studied; modular and imitation learning agents were not evaluated.
- Visual realism metrics reflect the dataset-renderer combination using a simple rasterization renderer in Habitat.
- Fine-tuned results show the gap largely closes after finetuning on the target dataset.
Experimental setup
Datasets: HSSD-200 ProcTHOR-10K iTHOR HM3DSem MP3D Gibson
Benchmarks: HM3DSem ObjectNav val MP3D ObjectNav val iTHOR val ProcTHOR val HSSD val
Metrics: success rate SPL FID KID object co-occurrence similarity navigable area
Baselines: ProcTHOR-10K agents iTHOR agents HM3DSem-trained agents MP3D-trained agents
Builds on: DDPPO ProcTHOR Habitat Khandelwal et al. ObjectNav architecture
Closest papers by meaning
- Image2Sim: Scaling Embodied Navigation via Generative Neural Simulator (NeurIPS 2026 · similarity 0.90)
- InternScenes: A Large-scale Simulatable Indoor Scene Dataset with Realistic Layouts (NeurIPS 2025 · similarity 0.88)
- SceneWeaver: All-in-One 3D Scene Synthesis with an Extensible and Self-Reflective Agent (NeurIPS 2025 · similarity 0.87)
- AmaraSpatial-10K: A Spatially and Semantically Aligned 3D Dataset for Spatial Computing and Embodied AI (NeurIPS 2026 · similarity 0.87)
- Learning to Navigate Efficiently and Precisely in Real Environments (CVPR 2024 · similarity 0.87)
- RoomTour3D: Geometry-Aware Video-Instruction Tuning for Embodied Navigation (CVPR 2025 · similarity 0.86)
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.