wuvin
Always take risks!
Toggle navigation
wuvin
主页
实验室的搬砖生活
机器学习
公开的学术内容
公开的其他内容
About Me
归档
标签
友情链接
ZYQN
ihopenot
enigma_aw
hzwer
杨宗翰
Mutual Mean-Teaching: Pseudo Label Refinery for Unsupervised Domain Adaptation on Person Re-identification
? Domain Adaptation ?
2020-06-16 19:29:37
602
0
0
wuvin
? Domain Adaptation ?
# Mutual Mean-Teaching: Pseudo Label Refinery for Unsupervised Domain Adaptation on Person Re-identification ## Info * Conference: ICLR 2020 * Cites: 6 * Github Stars: 211 * Github Solved/Issue: 22/26 * Author: ![title](https://leanote.com/api/file/getImage?fileId=5ee8ad95ab64416c2000181c) ## Main Idea * Method for Unsupervised DA on Person Re-ID. [Person re-identification (re-ID) aims at identifying the same persons’ images across different cameras. However, domain diversities between different datasets pose an evident challenge for adapting the re-ID model trained on one dataset to another one.] * Traditional methods usually use $Clusters + Pseudo Lable$, however a wrong pseudo lable usually have great bad effect on performance. * So, they solve this with $Mutual Learning + Moving Average Model$. The moving part is design to avoid unstable training. * ![title](https://leanote.com/api/file/getImage?fileId=5ee8b224ab64416a26001800) * mAP 提高了**十来个点** (有两个模型+做这个数据集的人少+Mutual Learning Buff) * Ablation studies, 可以看出模型平均非常有用 * ![title](https://leanote.com/api/file/getImage?fileId=5ee8b3b3ab64416c2000185d) * 想起之前一篇UDA,用 moving average 标pseudo lable,效果非常好。应该是 Moving Average 在 pseudo lable 相关的地方都很有用。
上一篇:
Switchable Whitening for Deep Representation Learning
下一篇:
paper reading 可解释性
0
赞
602 人读过
新浪微博
微信
腾讯微博
QQ空间
人人网
提交评论
立即登录
, 发表评论.
没有帐号?
立即注册
0
条评论
More...
文档导航
没有帐号? 立即注册