AI Paper Atlas

Research map / Architectures and efficiency / CNN design

BA-Net: Bridge Attention for Deep Convolutional Neural Networks

Yue Zhao, Junzhou Chen, Zirui Zhang, Ronghui Zhang

ECCV 2022 · Architectures and efficiency › CNN design · approach: Efficient CNN scaling and dynamic inference

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
bridge conv1 and conv2 outputs of the current block into the attention layer (ablation)SE ResNet-50 · ImageNet-1Ktop-1 accuracy (%): 78.14 -> 78.85Improved
bridge the previous block's attention weights instead of the current block's convs (ablation)BA ResNet-50 · ImageNet-1Ktop-1 accuracy (%): 78.85 -> 78.41Hurt
bridge previous block conv3 instead of current block convs (ablation)BA ResNet-50 · ImageNet-1Ktop-1 accuracy (%): 78.85 -> 78.49Hurt
bridge previous block end output instead of current block convs (ablation)BA ResNet-50 · ImageNet-1Ktop-1 accuracy (%): 78.85 -> 78.54Hurt
bridge only current conv1 rather than conv1 and conv2 (ablation)BA ResNet-50 · ImageNet-1Ktop-1 accuracy (%): 78.85 -> 78.78Hurt
bridge only current conv2 rather than conv1 and conv2 (ablation)BA ResNet-50 · ImageNet-1Ktop-1 accuracy (%): 78.85 -> 78.77Hurt
add bridge attention (ablation)ResNeXt-50 · ImageNet-1Ktop-1 accuracy (%): 78.77 -> 79.58Improved
add bridge attention (ablation)EfficientNet-b0 · ImageNet-1Ktop-1 accuracy (%): 70.11 -> 71.70Improved
add bridge attention (ablation)MobileNetv3-small · ImageNet-1Ktop-1 accuracy (%): 65.87 -> 65.85No clear effect
add bridge attention (ablation)MobileNetv3-large · ImageNet-1Ktop-1 accuracy (%): 73.29 -> 73.50Improved
replace SENet attention with BA-Net (ablation)ResNet-101 · ImageNet-1Ktop-1 accuracy (%): 79.41 -> 80.03Improved
replace SENet attention with BA-Net (ablation)ResNet-50 Faster R-CNN · COCO2017box mAP: 37.7 -> 39.5Improved

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