AI Paper Atlas

Research map / Detection and segmentation / CNN features

Conformer: Local Features Coupling Global Representations for Visual Recognition

Zhiliang Peng, Wei Huang, Shanzhi Gu, Lingxi Xie, Yaowei Wang, Jianbin Jiao, Qixiang Ye

ICCV 2021 · Detection and segmentation › CNN features · approach: Convolution-transformer backbones · 934 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
replace DeiT-B with Conformer-BDeiT-B · ImageNet-1k valtop-1 accuracy: 81.8 -> 84.1Improved
use Conformer-S/32 backbone instead of ResNet-101 in FPNResNet-101 FPN · MSCOCO minivalAP bbox: 39.4 -> 43.1Improved
use Conformer-S/32 backbone instead of ResNet-101 in Mask R-CNNResNet-101 Mask R-CNN · MSCOCO minivalAP segm: 36.1 -> 39.7Improved
add learnable positional embeddings to DeiT-S (ablation)DeiT-S · ImageNet-1k valtop-1 accuracy: 77.4 -> 79.8Improved
add learnable positional embeddings to Conformer-S (ablation)Conformer-S · ImageNet-1k valtop-1 accuracy: 83.4 -> 83.5No clear effect
fuse every block instead of every 4 blocks (ablation)Conformer-S · ImageNet-1k valtop-1 accuracy: 82.2 -> 83.4Improved
fuse every block instead of every 2 blocks (ablation)Conformer-S · ImageNet-1k valtop-1 accuracy: 82.9 -> 83.4Improved
use attention-based sampling instead of avg pooling (ablation)Conformer-S · ImageNet-1k valtop-1 accuracy: 83.4 -> 83.3No clear effect
use avg pooling instead of max pooling for sampling (ablation)Conformer-S · ImageNet-1k valtop-1 accuracy: 83.3 -> 83.4No clear effect
couple branches with FCU instead of ensembling DeiT-S and ResNet-101 (ablation)DeiT-S + ResNet-101 · ImageNet-1k valtop-1 accuracy: 81.8 -> 83.4Improved
use dual structure instead of serial ResNet-50d & DeiT-S hybrid (ablation)ResNet-50d & DeiT-S · ImageNet-1k valtop-1 accuracy: 80.4 -> 81.9Improved
increase CNN branch channels from 64 to 256 (ablation)Conformer (E=384) · ImageNet-1k valtop-1 accuracy: 81.3 -> 83.4Improved

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