Research map / Label-efficient and robust learning / Domain adaptation and generalization
A New Benchmark: On the Utility of Synthetic Data With Blender for Bare Supervised Learning and Downstream Domain Adaptation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Train on non-repetitive rendered samples instead of repeating a fixed dataset | ResNet-50 · SubVisDA-10 synthetic IID | accuracy: 94.86 -> 96.26 | Helped |
| Train on non-repetitive rendered samples instead of repeating a fixed dataset | ResNet-50 · SubVisDA-10 OOD | accuracy: 42.24 -> 42.82 | Helped |
| Increase augmentation strength (ablation) | Non-repetitive ResNet-50 training · SubVisDA-10 | OOD accuracy: 25.04 -> 27.47 -> 42.82 | Helped |
| Increase augmentation strength (ablation) | Non-repetitive ResNet-50 training · SubVisDA-10 synthetic IID | accuracy: 98.19 -> 98.10 -> 96.26 | Hurt |
| Add variation in object scale rather than fixing one scale (ablation) | ResNet-50 synthetic supervised learning · SubVisDA-10 synthetic IID | accuracy: 68.77 / 80.80 / 77.61 -> 87.12 | Helped |
| Randomize material texture instead of fixing one texture (ablation) | ResNet-50 synthetic supervised learning · SubVisDA-10 synthetic IID | accuracy: 79.58 / 50.29 / 50.35 / 68.62 -> 87.63 | Helped |
| Randomize camera viewpoint instead of fixing it (ablation) | ResNet-50 synthetic supervised learning · SubVisDA-10 synthetic IID | accuracy: 24.60 / 27.21 / 32.82 / 33.79 -> 87.63 | Helped |
| Include randomized backgrounds instead of removing backgrounds (ablation) | ResNet-50 synthetic supervised learning · SubVisDA-10 synthetic IID | accuracy: 17.68 -> 87.63 | Helped |
| Remove backgrounds from training images (ablation) | ResNet-50 synthetic supervised learning · SubVisDA-10 synthetic IID without background | accuracy: 78.55 -> 94.75 | Helped |
| Pre-train on SynSL 4.8M rather than SynSL 200K (ablation) | MCD domain adaptation · SubVisDA-10 target | accuracy: 55.06 -> 65.69 | Helped |
| Add a small set of real target-class data to synthetic pre-training (ablation) | Domain adaptation pre-training · SubVisDA-10 target | accuracy: Ours: 55.23 (MCD); Ours+SubImageNet: 60.90 (MCD) | Helped |
| Add synthetic target-class data to ImageNet-990 pre-training (ablation) | Domain adaptation pre-training · SubVisDA-10 target | accuracy: 58.84 -> 60.52 (MCD) | Helped |
Closest papers by meaning
- Real-Fake: Effective Training Data Synthesis Through Distribution Matching (ICLR 2024 · similarity 0.88)
- The Unmet Promise of Synthetic Training Images: Using Retrieved Real Images Performs Better (NeurIPS 2024 · similarity 0.88)
- Scaling Laws of Synthetic Images for Model Training ... for Now (CVPR 2024 · similarity 0.87)
- BLURD: Benchmarking and Learning using a Unified Rendering and Diffusion Model (NeurIPS 2024 · similarity 0.87)
- AUGCAL: Improving Sim2Real Adaptation by Uncertainty Calibration on Augmented Synthetic Images (ICLR 2024 · similarity 0.86)
- Provably Improving Generalization of Few-shot models with Synthetic Data (ICML 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.