AI Paper Atlas

Research map / Architectures and efficiency / Attention and transformers

Removing Rows and Columns of Tokens in Vision Transformer enables Faster Dense Prediction without Retraining

Diwei Su, cheng fei, Jianxu Luo

ECCV 2024 · Architectures and efficiency › Attention and transformers · approach: Efficient and robust vision transformers · 1 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
apply Token Adapter to Segmenter at best-accuracy settingSegmenter ViT-L/16 · ADE20KmIoU: 51.82 -> 52.02 with FPS 10.68 -> 15.09Improved
replace clustering (Expediting) with matching (Token Adapter) at max speedupSegmenter ViT-L/16 · ADE20KmIoU: 43.86 -> 45.73Improved
replace ToMe with Token Adapter at max speedupSegmenter ViT-L/16 · CityscapesFPS: 13.12 -> 19.38Improved
apply Token Adapter instead of Expediting at early MaskDINO stagesMaskDINO Swin-L · COCO panopticPQ: 30.7 -> 55.91 (stage 1)Improved
apply Token Adapter at stage 4 of MaskDINOMaskDINO Swin-L · COCO panopticPQ: 58.3 -> 58.42No clear effect
use Token Adapter instead of Token Merging at aggressive r=64 on SWAGSWAG ViT-L/16 · ImageNet-1k valaccuracy: 67.56 -> 85.34Improved
select rows/columns by feature importance scores instead of random (ablation)Token Adapter · ADE20KmIoU: 51.38 -> 52.02Improved
use feature scores instead of attention scores for selection (ablation)Token Adapter · ADE20KmIoU: 51.13 -> 52.02Improved
use cosine distance instead of other metrics for matching (ablation)Token Adapter · ADE20KmIoU: 51.80 -> 52.02 (dot)Improved
merge by averaging instead of pooling or bilinear resize (ablation)Token Adapter · ADE20KmIoU: 48.15 -> 52.02 (max pool)Improved
merge normal tokens instead of just keeping representative and unique tokens (ablation)Token Adapter · ADE20KmIoU: 50.67 -> 52.02Improved
use threshold matching in ejector instead of one-to-one max (ablation)Token Adapter · ADE20KmIoU: 51.93 -> 52.02Improved

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