AI Paper Atlas

Research map / Architectures and efficiency / Attention and transformers

SG-Former: Self-guided Transformer with Evolving Token Reallocation

Sucheng Ren, Xingyi Yang, Songhua Liu, Xinchao Wang

ICCV 2023 · Architectures and efficiency › Attention and transformers · approach: Efficient and robust vision transformers · 74 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
use hybrid-scale attention instead of single-scale local attention (ablation)SG-Former · ImageNet-1K valTop-1 accuracy (%): 81.6 -> 82.6Improved
use hybrid-scale attention instead of single-scale global attention (ablation)SG-Former · ImageNet-1K valTop-1 accuracy (%): 82.0 -> 82.6Improved
use learned significance map for token reallocation instead of uniform manual map (ablation)SG-Former · ImageNet-1K valTop-1 accuracy (%): 82.3 -> 83.2Improved
use hybrid-scale significance instead of local-only significance (ablation)SG-Former · ImageNet-1K valTop-1 accuracy (%): 82.9 -> 83.2Improved
use global-only significance instead of manual uniform map (ablation)SG-Former · ImageNet-1K valTop-1 accuracy (%): 82.3 -> 82.4No clear effect
SG-Former-S versus Swin-T at 224 resolutionbackbone · ImageNet-1K valTop-1 accuracy (%): 81.2 -> 83.2Improved
SG-Former-B versus CSWin-B at 224 resolutionbackbone · ImageNet-1K valTop-1 accuracy (%): 84.2 -> 84.7Improved
SG-Former-S versus CSWin-T on COCO Mask R-CNN 1xbackbone · COCO Mask R-CNN 1x (small)box AP: 46.7 -> 47.4Improved
SG-Former-B versus Swin-B on ADE20K UperNetbackbone · ADE20K UperNet 160k (base)mIoU: 48.1 -> 52.0Improved

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