Research map / Architectures and efficiency / Attention and transformers
Rotary Position Embedding for Vision Transformer
Byeongho Heo, Song Park, Dongyoon Han, Sangdoo Yun
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 |
|---|---|---|---|
| replace APE with RoPE-Mixed in ViT-B | ViT-B · ImageNet-1k val, 512x512 test | top-1 accuracy: 80.5 -> 82.9 | Improved |
| use mixed learnable-frequency RoPE instead of axial frequency RoPE (ablation) | ViT-B · ImageNet-1k val, multi-resolution | top-1 accuracy: 82.0 -> 82.9 at 512x512 | Improved |
| make axial RoPE frequencies learnable without mixing axes (ablation) | ViT-S RoPE-Axial · ImageNet-1k val, ViT-S | top-1 accuracy: -0.2 at 224, +1.3 at 384, +1.7 at 512 | Mixed |
| add APE on top of RoPE-Mixed in ViT-B (ablation) | ViT-B RoPE-Mixed · ImageNet-1k val, 112 vs 512 test | top-1 accuracy gain over APE: 1.1 -> 2.2 at 112; 2.4 -> 1.9 at 512 | Mixed |
| add RPB on top of RoPE-Mixed in Swin-B (ablation) | Swin-B RoPE-Mixed · ImageNet-1k val, 384x384 | gain over RPB: 0.9 -> 0.2 | Hurt |
| replace APE with RoPE-Mixed in DINO-ViTDet backbone | ViT-B · MS-COCO detection | box AP: 49.4 -> 51.2 | Improved |
| replace RPB with RoPE-Mixed in DINO-Swin backbone | Swin-T · MS-COCO detection | box AP: 51.3 -> 51.8 | Improved |
| replace APE with RoPE-Mixed + APE in UperNet ViT-B | ViT-B · ADE20k UperNet single-scale | mIoU: 47.7 -> 50.0 | Improved |
| add RPB to axial RoPE in Swin-B Mask2Former (ablation) | Swin-B RoPE-Axial · ADE20k Swin-Mask2Former | mIoU: 52.0 -> 50.0 | Hurt |
| use RoPE-Mixed instead of ResFormer-S-224 at small resolution | ViT-S · ImageNet-1k val, 128x128 | top-1 accuracy: 71.4 -> 70.6 | Hurt |
| use RoPE-Mixed instead of ResFormer-S-224 at large resolution | ViT-S · ImageNet-1k val, 512x512 | top-1 accuracy: 77.7 -> 79.1 | Improved |
Closest papers by meaning
- ComRoPE: Scalable and Robust Rotary Position Embedding Parameterized by Trainable Commuting Angle Matrices (CVPR 2025 · similarity 0.88)
- A Circular Argument: Does RoPE need to be Equivariant for Vision? (NeurIPS 2025 · similarity 0.88)
- Circle-RoPE: Cone-like Decoupled Rotary Positional Embedding for Vision-Language Models (ICML 2026 · similarity 0.88)
- VideoRoPE: What Makes for Good Video Rotary Position Embedding? (ICML 2025 · similarity 0.87)
- LieRE: Lie Rotational Positional Encodings (ICML 2025 · similarity 0.86)
- Maximizing the Position Embedding for Vision Transformers with Global Average Pooling (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.
