AI Paper Atlas

Research map / Architectures and efficiency / Attention and transformers

When Token Compression Breaks: Structural Pruning vs. Token Reduction for Robust ViT Segmentation under High Compression

Tien-Phat Nguyen, Ngai-Man Cheung

ECCV 2026 · Architectures and efficiency › Attention and transformers · approach: Vision transformer pruning · 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
mild token compression with CTSDeiT-B Segmenter · ADE20K valmIoU clean/noise: 47.22/35.58 -> 46.03/34.76No clear effect
mild token compression with CTSDeiT-B Segmenter · Cityscapes valmIoU clean/noise: 78.25/63.26 -> 78.15/63.07No clear effect
aggressive ToMe merging (3.7x FLOPs reduction)DeiT-B Segmenter · ADE20K valmIoU clean/noise: 47.22/35.58 -> 23.94/16.86Hurt
aggressive ToMe mergingDeiT-B Segmenter · Cityscapes valmIoU clean/noise: 78.25/63.26 -> 64.50/50.56Hurt
aggressive segmentation-aware token compression (ALGM)DeiT-B Segmenter · ADE20K valmIoU clean: 47.22 -> 42.85Hurt
structural pruning (NViT-S) at 3.8x FLOPs reductionDeiT-B Segmenter · ADE20K valmIoU clean: 47.22 -> 45.30Hurt
mild structural pruning (NViT) at 1.4xDeiT-B Segmenter · ADE20K valmIoU clean: 47.22 -> 47.93Improved
prune-then-merge instead of ToMe alone at high compression (ablation)NViT-B+ToMe · ADE20K valmIoU clean: 23.94 -> 45.71Improved
prune-then-merge instead of NViT-S alone at high compression (ablation)NViT-B+ToMe · ADE20K valmIoU noise: 32.86 -> 34.33Improved
prune-then-merge instead of NViT-S alone (ablation)NViT-B+ToMe · Cityscapes valmIoU clean: 76.76 -> 77.52Improved
prune-then-merge instead of CTS at high compression (ablation)NViT-B+ToMe · Cityscapes-CmIoU noise: 60.99 -> 60.71Hurt
prune-then-merge vs very aggressive pruning at matched FPS (ablation)NViT-B+ToMe · ADE20K val FPS-matchedmIoU clean: 41.01 -> 45.17Improved

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