Research map / Detection and segmentation / CNN features
MPViT: Multi-Path Vision Transformer for Dense Prediction
Check 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 |
|---|---|---|---|
| Replace a single path with dual paths (ablation) | MPViT-XS · ImageNet-1K and COCO | Top-1 accuracy and Mask R-CNN box AP: 78.9 -> 80.7; 40.2 -> 42.6 | Helped |
| Expand to triple paths while reducing the number of layers (ablation) | MPViT-XS · ImageNet-1K and COCO | Top-1 accuracy and Mask R-CNN box AP: 80.7 -> 79.8; 42.6 -> 41.4 | Hurt |
| Expand to triple paths while reducing embedding dimensions rather than layers (ablation) | MPViT-XS · ImageNet-1K and COCO | Top-1 accuracy and Mask R-CNN box AP: 80.7 -> 80.5; 42.6 -> 43.0 | Mixed |
| Expand from triple paths to quad paths (ablation) | MPViT-XS · ImageNet-1K and COCO | Top-1 accuracy and Mask R-CNN box AP: 80.5 -> 80.5; 43.0 -> 42.4 | Mixed |
| Use consecutive 3×3 patch-embedding convolutions instead of parallel 3×3, 5×5, and 7×7 convolutions (ablation) | MPViT-XS · ImageNet-1K and COCO | Top-1 accuracy and Mask R-CNN box AP: 80.3 -> 80.5; 41.7 -> 43.0 | Helped |
| Aggregate local and global features by summation (ablation) | MPViT-XS GLI · ImageNet-1K | Top-1 accuracy: 80.5 -> 80.3 | No clear effect |
| Aggregate local and global features by concatenation (ablation) | MPViT-XS GLI · ImageNet-1K and COCO | Top-1 accuracy and Mask R-CNN box AP: 80.5 -> 80.8; 43.0 -> 43.3 | Helped |
| Use MPViT-B instead of Focal-B | ImageNet-1K · ImageNet-1K | Top-1 accuracy: 83.8 -> 84.3 | Helped |
| Use MPViT-B instead of Focal-B | COCO, ADE20K · COCO Mask R-CNN box AP; ADE20K mIoU | box AP and mIoU: 49.0 -> 49.5; 49.0 -> 50.3 | Helped |
Closest papers by meaning
- AiluRus: A Scalable ViT Framework for Dense Prediction (NeurIPS 2023 · similarity 0.89)
- Multi-Scale High-Resolution Vision Transformer for Semantic Segmentation (CVPR 2022 · similarity 0.89)
- ViT-CoMer: Vision Transformer with Convolutional Multi-scale Feature Interaction for Dense Predictions (CVPR 2024 · similarity 0.88)
- Auto-scaling Vision Transformers without Training (ICLR 2022 · similarity 0.88)
- Your ViT is Secretly an Image Segmentation Model (CVPR 2025 · similarity 0.87)
- Data-independent Module-aware Pruning for Hierarchical Vision Transformers (ICLR 2024 · similarity 0.86)
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.