Research map / Label-efficient and robust learning / Domain adaptation and generalization
Factors of Influence for Transfer Learning Across Diverse Appearance Domains and Task Types
Thomas Mensink, Jasper Uijlings, Alina Kuznetsova, Michael Gygli, Vittorio Ferrari
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 |
|---|---|---|---|
| start from ILSVRC'12 pre-training instead of training from scratch | HRNetV2 transfer models · full target training set, all targets | relative performance: improves performance by 5%-46% | Improved |
| add an intermediate best source task between ILSVRC'12 and the target | ILSVRC'12 pre-trained models · small target training set | share of targets with positive transfer: 85% positive, 67% very positive | Improved |
| add an intermediate best source task between ILSVRC'12 and the target | ILSVRC'12 pre-trained models · full target training set | share of targets with positive transfer: 56% positive, 19% very positive | Improved |
| use a source from the same image domain instead of a different domain (same task type) (ablation) | transfer chains · small target training set | share with positive transfer: 14% -> 69% | Improved |
| use a source of the same task type instead of another task type (different domain) (ablation) | transfer chains · small target training set | share with positive transfer: 5% -> 14% | Improved |
| measure domain distance by target-to-closest-source inclusion rather than source-to-closest-target (ablation) | domain distance measure · small source small target semantic segmentation | Kendall tau with transfer gain: 0.06 -> 0.42 | Improved |
| compare domain distance to source size as predictor of transfer gain (ablation) | small target transfer experiments · semantic segmentation | Kendall tau with transfer gain: 0.12 -> 0.40 | Improved |
| use a larger source than the target dataset (within-domain, within-task) (ablation) | transfer chains · small target training set | share with positive transfer: 5% -> 60% | Improved |
| use the largest within-domain source instead of the multi-source model (ablation) | semantic segmentation transfer · small target training set, 10 targets | transfer gain: 7 out of 10 significantly better | Mixed |
| use a full target training set instead of a small one (ablation) | transfer chains · all source-target pairs | share with positive transfer: 21% -> 14% | Hurt |
| use segmentation source for object detection on the same dataset | cross-task transfer · small target training set | share with positive transfer: 100% positive | Improved |
| use keypoint detection source for segmentation on the same dataset | cross-task transfer · small target training set | share with negative transfer: 100% negative | Hurt |
Closest papers by meaning
- Taskonomy: Disentangling Task Transfer Learning (CVPR 2018 · similarity 0.89)
- Learning Across Tasks and Domains (ICCV 2019 · similarity 0.89)
- Understanding the Transferability of Representations via Task-Relatedness (NeurIPS 2024 · similarity 0.89)
- Which Model To Transfer? Finding the Needle in the Growing Haystack (CVPR 2022 · similarity 0.89)
- Knowledge Transfer from Vision Foundation Models for Efficient Training of Small Task-specific Models (ICML 2024 · similarity 0.88)
- On the Surprising Effectiveness of Attention Transfer for Vision Transformers (NeurIPS 2024 · 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.
