Research map / Vision-language and multimodal / Human and embodied modeling
V-PETL Bench: A Unified Visual Parameter-Efficient Transfer Learning Benchmark
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Builds V-PETL Bench, a unified benchmark of 30 image, video, and dense-prediction datasets for evaluating 25 parameter-efficient transfer learning methods on ViT and Swin backbones, with an open-source codebase and a new PPT metric.
Problem
Comparing parameter-efficient transfer learning (PETL) algorithms in computer vision is inconvenient: hyperparameters are often unreleased, baselines are underestimated, and existing benchmarks cover only plain image recognition.
Approach
A unified, open-source benchmark spanning image recognition, video action recognition, and dense prediction, with per-algorithm hyperparameter search and a single Performance-Parameter Trade-off (PPT) metric that combines accuracy with trainable parameter count.
Key results
- On FGVC, GPS (91.78 mean) and SNF (90.74) lead; several PETL methods outperform full fine-tuning (88.54 mean).
- On VTAB, DTL achieves the best PPT (0.75) with 74.58 mean accuracy using 0.04M trainable parameters, versus 65.57 for full fine-tuning.
- On ADE20K, full fine-tuning (52.10 mIoU) outperforms all nine evaluated PETL methods, while on PASCAL VOC all PETL methods beat full fine-tuning.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use GPS instead of Full fine-tuning | ViT-B/16 · FGVC mean | accuracy: 88.54 -> 91.78 | Helped |
| Use DTL instead of Full fine-tuning | ViT-B/16 · VTAB mean | accuracy: 65.57 -> 74.58 | Helped |
| Use AdaptFormer instead of Full fine-tuning | ViT-B (VideoMAE) · SSv2 | Top-1 accuracy: 53.97 % -> 59.02 % | Helped |
| Use BAPAT instead of Full fine-tuning | Video Swin-B · HMDB51 | Top-1 accuracy: 68.07 % -> 71.93 % | Helped |
| Use Mona instead of Full fine-tuning | Swin-B · COCO | APBox: 51.9 % -> 53.4 % | Helped |
| Use LN TUNE instead of Full fine-tuning | Swin-L · PASCAL VOC | APBox: 83.5 % -> 85.8 % | Helped |
| Use LN TUNE instead of Full fine-tuning | Swin-L · ADE20K | mIoU: 52.10 % -> 47.98 % | Hurt |
| Apply PETL algorithms to limited-data image recognition instead of Full fine-tuning | ViT-B/16 · VTAB | accuracy: 65.57 -> 75.99 | Helped |
| Apply PETL algorithms directly from NLP without vision-specific design | Vision tasks · FGVC | accuracy | Hurt |
Surprising findings
- On ADE20K, full fine-tuning outperforms all PETL methods, while on PASCAL VOC every PETL method surpasses it.
- Directly applying NLP-origin PETL methods (Adapter, LoRA, BitFit) to vision yields degraded performance versus vision-specific designs.
Limitations
- Hyperparameters are searched per algorithm on validation splits, and results depend on the chosen backbones (ViT-B/L/H, Swin) and pretraining data.
- Benchmark covers only image, video, and dense prediction; multimodal and generative models are deferred to future work.
- Different tables evaluate different subsets of algorithms (9 to 18), so coverage is not uniform across tasks.
Experimental setup
Datasets: CUB-200-2011 NABirds Oxford Flowers Stanford Dogs Stanford Cars VTAB-1k Kinetics-400 SSv2 HMDB51 MS COCO ADE20K PASCAL VOC
Benchmarks: CUB-200-2011 NABirds Oxford Flowers Stanford Dogs Stanford Cars VTAB-1k SSv2 HMDB51 MS COCO PASCAL VOC ADE20K
Metrics: top-1 accuracy mIoU AP box AP mask trainable parameters PPT GPU memory
Baselines: Full fine-tuning Linear probing Frozen Adapter LoRA AdaptFormer VPT-Shallow VPT-Deep BitFit SSF
Builds on: Adapter LoRA VPT AdaptFormer VideoMAE
Closest papers by meaning
- DTL: Disentangled Transfer Learning for Visual Recognition (AAAI 2024 · similarity 0.88)
- UniPT: Universal Parallel Tuning for Transfer Learning with Efficient Parameter and Memory (CVPR 2024 · similarity 0.84)
- GPLQ: A General, Practical, and Lightning QAT Method for Vision Transformers (NeurIPS 2025 · similarity 0.84)
- VT-Bench: A Unified Benchmark for Visual-Tabular Multi-Modal Learning (ICML 2026 · similarity 0.84)
- Benchmarking and Enhancing VLM for Compressed Image Understanding (ICML 2026 · similarity 0.84)
- Time- Memory- and Parameter-Efficient Visual Adaptation (CVPR 2024 · similarity 0.83)
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.