Research map / Label-efficient and robust learning / Self-supervised and contrastive
Cross-Domain Self-Supervised Multi-Task Feature Learning Using Synthetic Imagery
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 |
|---|---|---|---|
| jointly predict edges, depth, and surface normals rather than train a single task (ablation) | synthetic-image feature learning · PASCAL VOC 2007 classification | classification score: 63.9, 61.9, 65.3 -> 65.6 | Helped |
| jointly predict edges, depth, and surface normals rather than train a single task (ablation) | synthetic-image feature learning · PASCAL VOC 2007 detection | detection score: 46.9, 48.9, 48.2 -> 51.3 | Helped |
| jointly predict edges, depth, and surface normals rather than train a single task (ablation) | synthetic-image feature learning · PASCAL VOC 2012 detection | detection score: 44.8, 45.8, 45.4 -> 47.2 | Helped |
| adapt features at conv5 (ablation) | three-task synthetic feature learning · PASCAL VOC 2007 classification | classification score: 65.6 -> 67.4 | Helped |
| adapt features at conv5 (ablation) | three-task synthetic feature learning · PASCAL VOC 2007 detection | detection score: 51.3 -> 52.0 | Helped |
| adapt features at conv5 (ablation) | three-task synthetic feature learning · PASCAL VOC 2012 detection | detection score: 47.2 -> 49.2 | Helped |
| perform domain adaptation at conv1 (ablation) | three-task synthetic feature learning · PASCAL VOC 2007 classification | classification score: 65.6 -> 61.9 | Hurt |
| adapt at conv6 instead of conv5 (ablation) | three-task synthetic feature learning · PASCAL VOC 2007 classification | classification score: 67.4 -> 66.9 | Hurt |
| fool the discriminator in both directions rather than only make synthetic features look real (ablation) | conv5 domain adaptation · PASCAL VOC 2007 classification | classification score: 67.4 -> 66.2 | Hurt |
| increase synthetic training data from 0.5M to 1.5M images (ablation) | three-task conv5-adapted model · PASCAL VOC 2007 classification | classification score: 67.4 -> 68.0 | Helped |
| increase synthetic training data from 0.5M to 1.5M images (ablation) | three-task conv5-adapted model · PASCAL VOC 2012 detection | detection score: 49.2 -> 50.0 | Helped |
| add real-image features to the generator update (ablation) | domain-adapted multi-task feature learning · transfer learning | transfer performance | No clear effect |
Closest papers by meaning
- Generative Modeling for Multi-task Visual Learning (ICML 2022 · similarity 0.88)
- Task2Sim: Towards Effective Pre-Training and Transfer From Synthetic Data (CVPR 2022 · similarity 0.87)
- Learning From Synthetic Data: Addressing Domain Shift for Semantic Segmentation (CVPR 2018 · similarity 0.87)
- A New Benchmark: On the Utility of Synthetic Data With Blender for Bare Supervised Learning and Downstream Domain Adaptation (CVPR 2023 · similarity 0.87)
- Visual Representation Learning over Latent Domains (ICLR 2022 · similarity 0.86)
- A General Representation-Based Approach to Multi-Source Domain Adaptation (ICML 2025 · similarity 0.85)
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.