Research map / Architectures and efficiency / Pruning and efficient tuning
DR-Tune: Improving Fine-tuning of Pretrained Visual Models by Distribution Regularization with Semantic Calibration
Nan Zhou, Jiaxin Chen, Di Huang
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 |
|---|---|---|---|
| add distribution regularization on the task head using pretrained features (ablation) | DR-Tune (CE-tuning base) · Caltech101 | top-1 accuracy (%): 91.93 -> 94.39 | Improved |
| add semantic calibration on top of distribution regularization (ablation) | DR-Tune · Caltech101 | top-1 accuracy (%): 94.39 -> 95.73 | Improved |
| add distribution regularization alone (no semantic calibration) on fine-grained data (ablation) | CE-tuning · Cars | top-1 accuracy (%): 88.45 -> 89.03 | Mixed |
| add semantic calibration on top of DR (ablation) | DR-Tune · Cars | top-1 accuracy (%): 89.03 -> 90.60 | Improved |
| add global rotation to SC (ablation) | DR-Tune SC module · Caltech101 | top-1 accuracy (%): 95.17 -> 95.73 | Improved |
| add class-level translation to SC (ablation) | DR-Tune SC module · Caltech101 | top-1 accuracy (%): 95.59 -> 95.73 | Improved |
| add confidence guided average for class centers (ablation) | DR-Tune SC module · Pets | top-1 accuracy (%): 89.69 -> 90.24 | Improved |
| class-level rotation instead of global rotation (ablation) | DR-Tune SC module · ImageNet20 | top-1 accuracy (%): 96.03 (GR) vs 95.82 (CLR) | No clear effect |
| apply DR-Tune instead of CE-tuning across pretraining strategies | ResNet-50 SimSiam · Caltech101 | top-1 accuracy (%): 82.28 -> 90.33 | Improved |
| reduce fine-tuning data to 10% of ImageNet20 | DR-Tune vs CE-tuning · ImageNet20 | top-1 accuracy drop: DR-Tune drop 3.3% vs CE-tuning drop 29.9% | Improved |
| use DR-Tune instead of knowledge distillation regularizers | ResNet-50 student · DTD | top-1 accuracy (%): 76.86 (best KD) -> 77.97 | Improved |
| use DR-Tune with ResNet-152 backbone | DR-Tune · DTD | top-1 accuracy (%): 70.16 -> 71.44 | Improved |
Later papers that compared with this method
1 later accepted paper reports DR-Tune as a baseline in their results tables; 0 of them report a better number on most of the results they share with it.
- TTE: Two Tokens Are Enough to Improve Parameter-Efficient Tuning (AAAI 2025 · better on 0 of 1 shared result)
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Towards Calibrated Robust Fine-Tuning of Vision-Language Models (NeurIPS 2024 · similarity 0.85)
- DA-VPT: Semantic-Guided Visual Prompt Tuning for Vision Transformers (CVPR 2025 · similarity 0.85)
- Low-Rank Rescaled Vision Transformer Fine-Tuning: A Residual Design Approach (CVPR 2024 · similarity 0.85)
- Learning Semantic Proxies from Visual Prompts for Parameter-Efficient Fine-Tuning in Deep Metric Learning (ICLR 2024 · similarity 0.85)
- Understanding and Mitigating Miscalibration in Prompt Tuning for Vision-Language Models (ICML 2025 · similarity 0.85)
- Proxy-FDA: Proxy-based Feature Distribution Alignment for Fine-tuning Vision Foundation Models without Forgetting (ICML 2025 · similarity 0.84)
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.
