




版权说明:本文档由用户提供并上传,收益归属内容提供方,若内容存在侵权,请进行举报或认领
文档简介
1、Training Region-based Object Detectors with Online Hard Example MiningAbhinav Shrivastava, Abhinav Gupta, Ross Girshick Focal Loss for Dense Object Detection Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, Piotr Dollar MentorNet: Regularizing Very Deep Neural Networks on Corrupted Labels Lu Ji
2、ang, Zhengyuan Zhou, Thomas Leung, Li-Jia Li, Li Fei-Fei Reporter: Fujin HeTraining Region-based Object Detectors with Online Hard Example MiningTraining Region-based Object Detectors with Online Hard Example MiningPART 01a large imbalance between the number of annotated objects and the number of ba
3、ckground examples , such as the deformable parts model (DPM) , this imbalance may beas extreme as 100,000 background examples to every oneobject. The recent trend towards object-proposal-based detectors mitigates this issue to an extent, but the imbalance ratio may still be high (e.g., 70:1) challen
4、geFor some period of time a fixed model is used to find new examples to add to the active training set.For some period of time the model is trained on the fixed active training set. ONEBenefitsIt removes the need for several heuristics and hyperparameters commonly used in region-based ConvNets .TWOI
5、t yields a consistent and significant boosts in mean average precision. THREEIts effectiveness increases as the training set es larger and more difficult, as emonstrated by results on the MS COCO dataset. Focal Loss for Dense Object Detection MentorNet: Regularizing Very Deep Neural Networks on Corr
6、upted Labels MentorNet: Regularizing Very Deep Neural Networks on Corrupted Labels PART 03TWOTHREEONEWe discover that deep CNNs trained on corr-upted labels can be improved by learning another network toweigh training examples. We propose an algorithm to optimize MentorNet with deep CNNs on big data
7、 and prove its conv-ergence under standard and mild assumptions. We empirically verify the proposed model on 4 datasets of both controlled and real-world noisy labels.ModelObjective FunctionD is training set, yi is the corresponding noisy label vector. Let gs denote the dis-Criminative function of a
8、 neural network called StuentendNet, L is a m-dimensional cloumn vector, denote the loss over m classes. v is a vector to represent the weight for i-th example, the function G is the explicit data regularizer(a technique adding a penalty term to the error function to prevent overfitting) ang lamda i
9、s the hyperparmeter, G determines the complexity of a weighting scheme, and imposes an (unnormalized) weight distribution over all training examples. AlgorithmIn the subroutine of minimizing w when fixing v, stochastic gradient descent often takes many steps before converging. This means that it can take a long time before moving past this single sub-step.More importantly, the subroutine of minimizing v
温馨提示
- 1. 本站所有资源如无特殊说明,都需要本地电脑安装OFFICE2007和PDF阅读器。图纸软件为CAD,CAXA,PROE,UG,SolidWorks等.压缩文件请下载最新的WinRAR软件解压。
- 2. 本站的文档不包含任何第三方提供的附件图纸等,如果需要附件,请联系上传者。文件的所有权益归上传用户所有。
- 3. 本站RAR压缩包中若带图纸,网页内容里面会有图纸预览,若没有图纸预览就没有图纸。
- 4. 未经权益所有人同意不得将文件中的内容挪作商业或盈利用途。
- 5. 人人文库网仅提供信息存储空间,仅对用户上传内容的表现方式做保护处理,对用户上传分享的文档内容本身不做任何修改或编辑,并不能对任何下载内容负责。
- 6. 下载文件中如有侵权或不适当内容,请与我们联系,我们立即纠正。
- 7. 本站不保证下载资源的准确性、安全性和完整性, 同时也不承担用户因使用这些下载资源对自己和他人造成任何形式的伤害或损失。
最新文档
- 环境监测仪表类型及应用考核试卷
- 封装材料特性考核试卷
- 派遣服务市场竞争力提升路径考核试卷
- 儿童性教育科学引导体系
- 婚内协议书范本
- 植物蓝染活动方案
- 楼梯装饰全年活动方案
- 油田职工比赛活动方案
- 油茶培训活动方案
- 河南废气处理活动方案
- FZ/T 10025-2022本色布技术要求规范
- YS/T 921-2013冰铜
- 刑法学(上册)马工程课件 第1章 刑法概说
- GB/T 9125.1-2020钢制管法兰连接用紧固件第1部分:PN系列
- GB/T 27770-2011病媒生物密度控制水平鼠类
- 2023年广西宾阳县昆仑投资集团有限公司招聘笔试题库及答案解析
- 社区社群团长招募书经典案例干货课件
- 12、施工现场“三级配电”配置规范-附电路图
- 新人教版七年级上册初中生物全册课时练(课后作业设计)
- 智能制造MES项目实施方案(注塑行业MES方案建议书)
- 四年级奥数讲义
评论
0/150
提交评论