AI Paper Atlas

Research map / Architectures and efficiency / Attention and transformers

Rethinking and Improving Relative Position Encoding for Vision Transformer

Kan Wu, Houwen Peng, Minghao Chen, Jianlong Fu, Hongyang Chao

ICCV 2021 · Architectures and efficiency › Attention and transformers · approach: Efficient and robust vision transformers · 427 citations (OpenAlex)

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 changedWhereMeasuredEffect
add learnable absolute position encoding to a model with no position encoding (ablation)DeiT-S · ImageNet-1k valtop-1 accuracy: 77.6 -> 79.9Improved
replace absolute position encoding with relative encoding on keys (ablation)DeiT-S · ImageNet-1k valtop-1 accuracy: 79.9 -> 80.9Improved
add absolute position encoding on top of relative encoding (ablation)DeiT-S · ImageNet-1k valtop-1 accuracy: 80.9 -> 80.9No clear effect
apply relative encoding to values instead of queries or keys (ablation)DeiT-S · ImageNet-1k valtop-1 accuracy: 80.9 -> 80.2Hurt
add relative encoding on queries and values to keys (ablation)DeiT-S · ImageNet-1k valtop-1 accuracy: 80.9 -> 81.3Improved
use directed Product instead of undirected Euclidean mapping (ablation)DeiT-S · ImageNet-1k valtop-1 accuracy: 80.4 -> 80.9Improved
use contextual instead of bias mode for Product method (ablation)DeiT-S · ImageNet-1k valtop-1 accuracy: 80.5 -> 80.9Improved
share bias-mode encoding across attention heads (ablation)DeiT-S · ImageNet-1k valtop-1 accuracy: 80.54 -> 80.05Hurt
share contextual-mode encoding across attention heads (ablation)DeiT-S · ImageNet-1k valtop-1 accuracy: 80.99 -> 80.89No clear effect
use piecewise instead of clip index function on classification (ablation)DeiT-S · ImageNet-1k valtop-1 accuracy: 80.1 -> 80.0No clear effect
use piecewise instead of clip index function (ablation)DETR-ResNet50 · COCO val, 150 epochsAP: 40.4 -> 40.8Improved
increase number of buckets up to 50 (ablation)DeiT-S contextual Product · ImageNet-1k valtop-1 accuracy: 79.9 -> 80.9Improved

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy