AI Paper Atlas

Research map / Detection and segmentation / CNN features

Conv2Former: A Simple Transformer-Style ConvNet for Visual Recognition

Qibin Hou, Cheng-Ze Lu, Ming‐Ming Cheng, Jiashi Feng

TPAMI 2024 · Detection and segmentation › CNN features · approach: Plain and high-resolution ViTs · 291 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
increase depthwise kernel size from 5x5 to 21x21 (ablation)Conv2Former-T · ImageNet-1ktop-1 accuracy: 82.8 -> 83.4Improved
increase depthwise kernel size from 5x5 to 21x21 (ablation)Conv2Former-B · ImageNet-1ktop-1 accuracy: 84.1 -> 84.5Improved
use 11x11 kernel instead of 7x7 (ablation)Conv2Former-B · ImageNet-1ktop-1 accuracy: 84.2 -> 84.4Improved
add Hadamard product modulation instead of element-wise sum (ablation)Conv2Former-T · ImageNet-1ktop-1 accuracy: 82.7 -> 83.2Improved
add Sigmoid after the convolutional weights A (ablation)Conv2Former-T · ImageNet-1ktop-1 accuracy: 83.2 -> 82.3Hurt
add L1 normalization after A (ablation)Conv2Former-T · ImageNet-1ktop-1 accuracy: 83.2 -> 82.8Hurt
linearly normalize A to (0, 1] (ablation)Conv2Former-T · ImageNet-1ktop-1 accuracy: 83.2 -> 82.2Hurt
use Layer Normalization with GELU (ablation)Conv2Former · ImageNet-1ktop-1 accuracy: +0.1%-0.2%Improved
use deeper stage configuration 3-3-12-3 instead of 2-2-8-2 for tiny models (ablation)Conv2Former · ImageNet-1ktop-1 accuracy: 81.5 -> 83.2Improved
use three convolutions for patch embedding instead of one (ablation)Conv2Former-IS · ImageNet-1ktop-1 accuracy: 81.2 -> 82.0Improved
use three convolutions for patch embedding instead of one (ablation)Conv2Former-IB · ImageNet-1ktop-1 accuracy: 82.7 -> 83.0Improved
replace ConvNeXt block with convolutional modulation blockConvNeXt-T · COCO Mask R-CNNbox AP: 46.2 -> 48.0Improved

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