Research map / Architectures and efficiency / Attention and transformers
ViT-Split: Unleashing the Power of Vision Foundation Models via Efficient Splitting Heads
Yifan Li, Xin Li, Tianqin Li, Wenbin He, Yu Kong, Liu Ren
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 prior head to frozen last-layer features (ablation) | ViT-Split · ADE20K (DINOv2-S) | mIoU: 44.3 -> 46.0 | Improved |
| add prior head to frozen last-layer features (ablation) | ViT-Split · ADE20K (DINOv2-B) | mIoU: 47.8 -> 51.4 | Improved |
| add task head (ablation) | ViT-Split · ADE20K (DINOv2-S) | mIoU: 44.3 -> 49.5 | Improved |
| add prior head to task head (ablation) | ViT-Split · ADE20K (DINOv2-S) | mIoU: 49.5 -> 50.4 | Improved |
| add prior head to task head (ablation) | ViT-Split · ADE20K (DINOv2-B) | mIoU: 53.2 -> 54.6 | Improved |
| add CNN fusion net instead of addition (ablation) | ViT-Split · ADE20K (DINOv2-S) | mIoU: 50.4 -> 51.6 | Improved |
| add CNN fusion net instead of addition (ablation) | ViT-Split · ADE20K (DINOv2-B) | mIoU: 54.6 -> 55.7 | Improved |
| use uniform layer selection instead of last few layers (ablation) | ViT-Split · ADE20K (DINOv2-L) | mIoU: 57.3 -> 58.2 | Improved |
| use sparse gate instead of uniform selection (ablation) | ViT-Split-L · Cityscapes 896 | mIoU: 85.8 -> 85.7 | No clear effect |
| use ViT-Split instead of end-to-end fine-tuned DINOv2 linear | DINOv2-L · Cityscapes 896 | mIoU (SS): 83.5 -> 85.8 | Improved |
| use ViT-Split instead of ViT-Adapter/ViT-CoMer on ADE20K 40K iterations | DINOv2-B · ADE20K 40K iters | mIoU: 49.9 -> 55.7 | Improved |
| use ViT-Split instead of ViT-CoMer on COCO | DINOv2-L · COCO Mask R-CNN | box AP: 53.4 -> 53.0 | Hurt |
Closest papers by meaning
- Swiss Army Knife: Synergizing Biases in Knowledge from Vision Foundation Models for Multi-Task Learning (ICLR 2025 · similarity 0.87)
- AVA-Bench: Atomic Visual Ability Benchmark for Vision Foundation Models (CVPR 2026 · similarity 0.86)
- Knowledge Transfer from Vision Foundation Models for Efficient Training of Small Task-specific Models (ICML 2024 · similarity 0.86)
- VMT-Adapter: Parameter-Efficient Transfer Learning for Multi-Task Dense Scene Understanding (AAAI 2024 · similarity 0.86)
- Unleashing the Power of Visual Foundation Models for Generalizable Semantic Segmentation (AAAI 2025 · similarity 0.86)
- VFM-Adapter: Adapting Visual Foundation Models for Dense Prediction with Dynamic Hybrid Operation Mapping (AAAI 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.
