Research map / Detection and segmentation / CNN features
Vision Transformer Adapter for Dense Predictions
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 the spatial prior module to ViT-S (ablation) | ViT-S with Mask R-CNN · COCO, 1× schedule | box AP: 40.2 -> 41.6; 1.4 AP^b | Helped |
| Use a spatial feature injector with cross-attention instead of direct addition (ablation) | ViT-S with spatial prior module and Mask R-CNN · COCO, 1× schedule | box AP: 41.6 -> 42.6; 1.0 AP^b | Helped |
| Add the multi-scale feature extractor (ablation) | ViT-S with Mask R-CNN · COCO, 1× schedule | box AP: 42.6 -> 44.7; 2.1 AP^b | Helped |
| Use 4 feature interactions instead of 1 (ablation) | ViT-Adapter-S with Mask R-CNN · COCO, 1× schedule | box AP: 43.2 -> 44.7 | Helped |
| Increase feature interactions from 4 to 6 (ablation) | ViT-Adapter-S with Mask R-CNN · COCO, 1× schedule | box AP: 44.7 -> 44.7 | No clear effect |
| Replace deformable attention with global attention (ablation) | ViT-Adapter-S · COCO detection | box AP: 44.7 -> 43.7 | Hurt |
| Replace deformable attention with CSwin attention (ablation) | ViT-Adapter-S · COCO detection | box AP: 44.7 -> 43.5 | Hurt |
| Replace deformable attention with Pale attention (ablation) | ViT-Adapter-S · COCO detection | box AP: 44.7 -> 44.2 | Hurt |
| Replace ImageNet-22K pre-training with multi-modal pre-training | ViT-Adapter-B with Mask R-CNN · COCO, 3×+MS | box AP: 50.5 -> 51.2; 0.7 AP^b | Helped |
| Replace ImageNet-22K pre-training with multi-modal pre-training | ViT-Adapter-L · ADE20K val, Semantic FPN | mIoU: 1.3 mIoU | Helped |
| Replace ImageNet-22K pre-training with multi-modal pre-training | ViT-Adapter-L · ADE20K val, UperNet | mIoU: 1.6 mIoU | Helped |
| Freeze the ViT backbone rather than full-tune (ablation) | ViT-Adapter-L · ADE20K | multi-scale mIoU: 53.7 -> 50.6 | Hurt |
Closest papers by meaning
- Memory Efficient Transformer Adapter for Dense Predictions (ICLR 2025 · similarity 0.94)
- ViT-CoMer: Vision Transformer with Convolutional Multi-scale Feature Interaction for Dense Predictions (CVPR 2024 · similarity 0.91)
- Your ViT is Secretly an Image Segmentation Model (CVPR 2025 · similarity 0.89)
- Dynamic Tuning Towards Parameter and Inference Efficiency for ViT Adaptation (NeurIPS 2024 · similarity 0.87)
- Probabilistic Conceptual Explainers: Trustworthy Conceptual Explanations for Vision Foundation Models (ICML 2024 · similarity 0.87)
- Linearly Decomposing and Recomposing Vision Transformers for Diverse-Scale Models (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.