AI Paper Atlas

Research map / Architectures and efficiency / CNN design

Convolution of Convolution: Let Kernels Spatially Collaborate

Rongzhen Zhao, Jian Li, Zhenzhi Wu

CVPR 2022 · Architectures and efficiency › CNN design · approach: Large and dynamic kernel design · 2 citations (OpenAlex)

Check 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 CoCImageNet classification backbones · ImageNet/valaccuracy: 70.01 -> 70.40; 76.25 -> 76.45; 74.03 -> 74.25; 75.82 -> 76.07; 69.59 -> 70.24; 66.34 -> 66.84Improved
combine CoC with orthogonalizationResNet18 · ImageNet/valaccuracy: 70.01 -> 71.20Improved
combine CoC with orthogonalizationResNet50 · ImageNet/valaccuracy: 76.25 -> 77.07Improved
add CoCRetinaNet-r50 · COCO2017/val bboxmAP: 36.3 -> 38.1Improved
add CoCMaskRCNN-hr18 · COCO2017/val maskmAP: 33.9 -> 35.6Improved
add CoCFCN-r50d8 · VOC0712/valmIoU: 66.97 -> 69.52Improved
add CoCHRNet-w18 · VOC0712/valmIoU: 72.13 -> 74.81Improved
add CoC while training from scratchFCN-r50d8 · VOC0712/valmIoU: 23.53 -> 34.25Improved
add CoC while training from scratchHRNet-w18 · VOC0712/valmIoU: 35.23 -> 41.39Improved
increase spatial association size beyond sa=2 (ablation)CoC · Tiny-ImageNetvalidation accuracyNo clear effect
increase super-kernel expansion too far (ablation)CoC · Tiny-ImageNetvalidation accuracyHurt
use more than one super layer (ablation)CoC · Tiny-ImageNetvalidation accuracyHurt

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