Research map / 3D vision / 3D generation
Meta-Sim2: Unsupervised Learning of Scene Structure for Synthetic Data Generation
Jeevan Devaranjan, Amlan Kar, Sanja Fidler
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| learn scene structure in addition to parameters (vs Meta-Sim learnt parameters only) (ablation) | Meta-Sim2 · KITTI-val AP@0.5 easy (tuned prior) | AP@0.5: 66.5 -> 67.0 | Improved |
| learn scene structure in addition to parameters | Meta-Sim2 · KITTI-train FID (tuned prior) | FID: 111.6 -> 99.7 | Improved |
| learn scene structure in addition to parameters | Meta-Sim2 · KITTI-train KID (tuned prior) | KID: 0.072 -> 0.054 | Improved |
| learn scene structure vs probabilistic grammar prior | Meta-Sim2 · KITTI-val AP@0.5 easy (tuned prior) | AP@0.5: 63.7 -> 67.0 | Improved |
| learn structure from a simple prior (prior parameters) | Meta-Sim2 · KITTI-val AP@0.5 easy (simple prior) | AP@0.5: 61.3 -> 63.2 | Improved |
| learn structure from a simple prior | Meta-Sim2 · KITTI-train FID (simple prior) | FID: 130.3 -> 100.0 | Improved |
| additionally learn parameters after learning structure (simple prior) | Meta-Sim2 · KITTI-val AP@0.5 easy (simple prior) | AP@0.5: 63.2 -> 65.2 | Improved |
| additionally learn parameters after learning structure (simple prior) | Meta-Sim2 · KITTI-train FID (simple prior) | FID: 100.0 -> 101.7 | No clear effect |
| start from simple prior instead of tuned prior | Meta-Sim2 · KITTI-train FID | FID: 99.7 -> 101.7 | No clear effect |
| replace per-sample KDE log-ratio reward with MMD batch score (ablation) | Meta-Sim2 · Multi-MNIST | digit distribution fit | Hurt |
Closest papers by meaning
- Compositional Scene Understanding through Inverse Generative Modeling (ICML 2025 · similarity 0.87)
- From Programs to Poses: Factored Real-World Scene Generation via Learned Program Libraries (NeurIPS 2025 · similarity 0.87)
- SimGen: Simulator-conditioned Driving Scene Generation (NeurIPS 2024 · similarity 0.86)
- InternScenes: A Large-scale Simulatable Indoor Scene Dataset with Realistic Layouts (NeurIPS 2025 · similarity 0.86)
- LT3SD: Latent Trees for 3D Scene Diffusion (CVPR 2025 · similarity 0.86)
- Disentangled 3D Scene Generation with Layout Learning (ICML 2024 · 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.
