Thursday, November 14, 2019

Res2Net

Res2Net

2019/11/12

前言:

Res2Net 可以說是 ResNet 一般化版本 ResNeXt 加上 SENet,是截至 2019 年底,CNN 表現最好的模型。某方面,它可說是把 CNN 基本的 conv1、identity mapping、cardinality 這三個觀念運用到極致!

-----


Fig. Res2Net and ResNeXt.

# Res2Net

-----


Fig. Res2Net and SENet.

# Res2Net

-----

References

◎ 論文

# Res2Net
Gao, Shang-Hua, et al. "Res2Net: A New Multi-scale Backbone Architecture." arXiv preprint arXiv:1904.01169 (2019).
https://arxiv.org/pdf/1904.01169.pdf  

-----

◎ 英文參考資料 

# 301 claps
Res2Net  New deep learning multi-scale architecture, for improved object detection with existing backbones
https://medium.com/@lessw/res2net-new-deep-learning-multi-scale-architecture-for-improved-object-detection-with-existing-de13095c9654 

Res2Net  A New Multi-scale Backbone Architecture – 程明明教授主页
https://mmcheng.net/res2net/

-----

◎ 簡體中文參考資料 

Res2Net:新型backbone网络,超越ResNet - 知乎
https://zhuanlan.zhihu.com/p/86331579 

超越ResNet:Res2Net;目标检测新网络:FCOS ;何恺明最新论文:RandWire-WS - 知乎
https://zhuanlan.zhihu.com/p/61407825

# 翻譯
Res2Net:新的深度学习多尺度结构,提升物体检测立竿见影
https://mp.weixin.qq.com/s?__biz=Mzg5ODAzMTkyMg==&mid=2247487222&idx=1&sn=2fb6814768c482ef93f08ecca4d9f75a&chksm=c06986abf71e0fbd58c6adf3de1af0929b33f1a3c642b1b94750f88de83e0ce4c4dc908b33da&scene=27#wechat_redirect

No comments: