Read the original document by opening this link in a new tab.
Table of Contents
1 Introduction
2 Related work
3 Revisiting ResNet, DenseNet and Higher Order RNN
4 Dual Path Networks
Summary
Dual Path Networks is a paper that introduces a novel architecture called the Dual Path Network (DPN) for image classification. It presents a new topology of connection paths internally, combining the benefits of Residual Network (ResNet) and Densely Convolutional Network (DenseNet). The DPN architecture enables effective feature re-usage and re-exploitation while maintaining high parameter efficiency, lower computational cost, and lower memory consumption. Extensive experiments on benchmark datasets demonstrate superior performance over existing state-of-the-art networks. The paper also discusses the relations between residual and densely connected networks, proposing a unified view on deep architecture design.