Research map / Label-efficient and robust learning / Domain adaptation and generalization
A Broad Study of Pre-training for Domain Generalization and Adaptation
Donghyun Kim, Kaihong Wang, Stan Sclaroff, Kate Saenko
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 |
|---|---|---|---|
| use Swin-T instead of ResNet-50 (both ImageNet-1K) | single-source DG · OH/CUB/WILD/DN avg | accuracy: 45.6 -> 49.3 | Improved |
| use Swin-S instead of ResNet-101 | single-source DG · OH/CUB/WILD/DN avg | accuracy: 48.5 -> 52.8 | Improved |
| use DeiT-S instead of ResNet-50 | single-source DG · CUB | accuracy: 42.3 -> 58.0 | Improved |
| use ConvNeXt-T instead of ResNet-50 | single-source DG · CUB | accuracy: 42.3 -> 42.5 | No clear effect |
| pretrain on ImageNet-22K instead of ImageNet-1K | ConvNeXt-B · CUB | accuracy: +19.8% | Improved |
| pretrain with ALBEF image-text data instead of ImageNet-supervised | ViT-B ALBEF · CUB | accuracy | Hurt |
| pretrain with ALBEF image-text data | ViT-B ALBEF · DomainNet | accuracy: 45.3 -> 50.3 | Improved |
| use MoCo self-supervised pretraining instead of supervised (ablation) | ResNet-50 · Office-Home | accuracy: 64.1 -> 53.7 | Hurt |
| use DINO self-supervised pretraining instead of supervised (ablation) | ResNet-50 · DomainNet | accuracy: 30.2 -> 23.1 | Hurt |
| add BEiT self-supervised learning to supervised pretraining (ablation) | ViT-L ImageNet-22K · CUB | accuracy: 76.5 -> 78.4 | Improved |
| add BEiT self-supervised learning to supervised pretraining (ablation) | ViT-L ImageNet-22K · Office-Home | accuracy: 82.3 -> 80.9 | Hurt |
| fine-tune the text encoder in ALBEF (ablation) | ALBEF · Office-Home | accuracy: 80.4 -> 79.3 | Hurt |
Closest papers by meaning
- Transferable Representation Learning with Deep Adaptation Networks (TPAMI 2018 · similarity 0.91)
- Generalized Domain Conditioned Adaptation Network (TPAMI 2021 · similarity 0.90)
- Image to Image Translation for Domain Adaptation (CVPR 2018 · similarity 0.90)
- Incorporating Pre-Training Data Matters in Unsupervised Domain Adaptation (TPAMI 2025 · similarity 0.90)
- Delving into Pre-training for Domain Transfer: A Broad Study of Pre-training for Domain Generalization and Domain Adaptation (IJCV 2026 · similarity 0.90)
- A General Representation-Based Approach to Multi-Source Domain Adaptation (ICML 2025 · similarity 0.89)
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.
