AI Paper Atlas

Research map / Architectures and efficiency / CNN design

DualNet: Learn Complementary Features for Image Recognition

Saihui Hou, Xu Liu, Zilei Wang

ICCV 2017 · Architectures and efficiency › CNN design · approach: Filter design and trained-filter analysis · 86 citations (OpenAlex)

Read the paper

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 auxiliary S2 Classifier during iterative training (ablation)DNI · CIFAR-100 (no aug)top-1 accuracy: 67.70% -> 69.09%Improved
add auxiliary classifiers and staged training instead of global training from scratch (ablation)DNI · CIFAR-100 (no aug)top-1 accuracy: 66.11% -> 69.76%Improved
add auxiliary classifiers and staged training instead of global finetuning from finetuned NIN (ablation)DNI · CIFAR-100 (no aug)top-1 accuracy: 68.49% -> 69.76%Improved
use DualNet instead of directly doubling feature maps per layer (ablation)NIN · CIFAR-100 (no aug)top-1 accuracy: 68.87% -> 69.76%Improved
directly double feature maps in NIN versus standard NIN (both from scratch) (ablation)NIN · CIFAR-100 (no aug)top-1 accuracy: 66.91% -> 68.87%Improved
add joint finetuning of the three classifiers after iterative training (ablation)DNI · CIFAR-100 (no aug)top-1 accuracy: 69.51% -> 69.76%Improved
add classifier averaging over three classifiers (ablation)DNI · CIFAR-100 (no aug)top-1 accuracy: 69.01% -> 69.51%Improved
add joint finetuning after iterative training (ablation)DNR · CIFAR-100 (ResNet-20, with aug)top-1 accuracy: 72.29% -> 72.43%Improved
use DualNet from ResNet-20 instead of single ResNet-20ResNet-20 · CIFAR-100 (ResNet-20, with aug)top-1 accuracy: 69.09% -> 72.43%Improved
use DualNet from VGGNet instead of VGGNetVGGNet-16 · Stanford Dogstop-1 accuracy: 74.11% -> 77.56%Improved
use DualNet from CaffeNet instead of CaffeNetCaffeNet · Stanford Dogstop-1 accuracy: 66.84% -> 67.94%Improved
use DualNet from VGGNet instead of VGGNetVGGNet-16 · UEC FOOD-100top-1 accuracy: 47.40% -> 49.19%Improved

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