AI Paper Atlas

Research map / Architectures and efficiency / Attention and transformers

EMOv2: Pushing 5M Vision Model Frontier

Jiangning Zhang, Teng Hu, Haoyang He, Zhucun Xue, Yabiao Wang, Chengjie Wang, Yong Liu, Xiangtai Li, Dacheng Tao

TPAMI 2025 · Architectures and efficiency › Attention and transformers · approach: Token merging and masked pretraining · 10 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 SEW-MHSA spanning attention to the block (ablation)EMOv2-5M · ImageNet-1K valTop-1: 73.5 -> 77.7Improved
add DW-Conv to the block (ablation)EMOv2-5M · ImageNet-1K valTop-1: 73.5 -> 78.1Improved
use both SEW-MHSA and DW-Conv (ablation)EMOv2-5M · ImageNet-1K valTop-1: 73.5 -> 79.4Improved
use spanning attention rather than neighborhood attention (ablation)EMOv2-5M · ImageNet-1K valTop-1: 78.8 -> 79.4Improved
use spanning attention rather than remote attention (ablation)EMOv2-5M · ImageNet-1K valTop-1: 79.0 -> 79.4Improved
add attention at all stages (S-1234) instead of stage 3-4 (ablation)EMOv2-5M · ImageNet-1K valTop-1: 79.4 -> 79.1Hurt
run DW-Conv and attention in cascade rather than parallel (ablation)EMOv1-5M · ImageNet-1K valTop-1: 78.1 -> 78.3Improved
switch order to DW-Conv first (ablation)EMOv1-5M / EMOv2-5M · ImageNet-1K valTop-1: -0.6 (EMOv1), -0.7 (EMOv2)Hurt
use kernel size 5 instead of 1 (ablation)EMOv2-5M · ImageNet-1K valTop-1: 78.6 -> 79.4Improved
replace DW-Conv with DCNv2 (ablation)EMOv2-5M · ImageNet-1K valTop-1: 79.4 -> 78.5Hurt
add iRMB to DeiT-TinyDeiT-Tiny · ImageNet-1K valTop-1: 72.2 -> 74.3Improved
add i2RMB to DeiT-TinyDeiT-Tiny · ImageNet-1K valTop-1: 72.2 -> 75.0Improved

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