当前位置:网站首页>[target tracking] | template update time context information (updatenet) "learning the model update for Siamese trackers"
[target tracking] | template update time context information (updatenet) "learning the model update for Siamese trackers"
2022-07-01 15:25:00 【rrr2】
Article title :《Learning the Model Update for Siamese Trackers》
Article address :[1] arXiv,[2] ICCV2019 openaccess
github Address :https://github.com/zhanglichao/updatenet
Update strategy
Use only the initial template of the first frame
Early Siamese In the tracker , Initialize the target template in the first frame , This template is used for tracking later . However, in the actual task, the change of target appearance is usually obvious , If the target template cannot be updated effectively , Soon the tracking will fail .
Linear update
To solve this problem , At present, some people have made some simple update strategies , Through the moving average method of fixed learning rate , Update the target template linearly :

We think this method has some disadvantages :
- For all the videos , It uses a fixed update rate . But in different cases , The need for updates is different , The factors to consider are complex . For example, camera motion blur , Dynamic changes at different times, etc .
- This update rate applies to all spatial dimensions of the template ( Including channel dimension ) It's also fixed on . But sometimes we only need to update a part of the template , For example, when occlusion occurs .
- The tracker cannot recover from the cumulative drift . As the template is updated , We lost T 0 The original information of , And the template of the first frame T0 That's what we first wanted to track .
- The formula (1) Just a very simple linear combination , But we need a flexible update mechanism , It's necessary to change the appearance when the goal is very complex . It may be better to use more complex combinatorial functions .
Adaptive update

The update algorithm will always consider the template of the first frame Provides highly reliable information , It enhances the robustness to the later tracking drift .


ref
translate
https://blog.csdn.net/PAN_Andy/article/details/103008079
https://blog.csdn.net/Yemiekai/article/details/123014839、
边栏推荐
- Zhang Chi's class: several types and differences of Six Sigma data
- Summary of empty string judgment in the project
- 使用swiper制作手机端轮播图
- 将ABAP On-Premises系统连接到中央检查系统以进行自定义代码迁移
- 【300+精选大厂面试题持续分享】大数据运维尖刀面试题专栏(三)
- 微信小程序01-底部导航栏设置
- DirectX repair tool v4.1 public beta! [easy to understand]
- Flink 系例 之 TableAPI & SQL 与 MYSQL 分组统计
- 入侵检测模型(An Intrusion-Detection Model)
- 【锁】Redis锁 处理并发 原子性
猜你喜欢

Skywalking 6.4 distributed link tracking usage notes

Basic use process of cmake

Junda technology - wechat cloud monitoring scheme for multiple precision air conditioners

cmake 基本使用过程

STM32F411 SPI2输出错误,PB15无脉冲调试记录【最后发现PB15与PB14短路】

Filter &(登录拦截)

异常检测中的浅层模型与深度学习模型综述(A Unifying Review of Deep and Shallow Anomaly Detection)

采集数据工具推荐,以及采集数据列表详细图解流程

微服务追踪SQL(支持Isto管控下的gorm查询追踪)

竣达技术丨多台精密空调微信云监控方案
随机推荐
Implementation of wechat web page subscription message
[one day learning awk] conditions and cycles
Redis installation and setting up SSDB master-slave environment under Ubuntu 14.04
Go zero actual combat demo (I)
MySQL 服务正在启动 MySQL 服务无法启动解决途径
k8s部署redis哨兵的实现
22-06-26周总结
SQL常用的四个排序函数梳理
Flink 系例 之 TableAPI & SQL 与 Kafka 消息获取
What is the relationship between network speed, broadband, bandwidth and traffic?
Returning to the top of the list, the ID is still weak
[advanced ROS] lesson 5 TF coordinate transformation in ROS
[antenna] [3] some shortcut keys of CST
[one day learning awk] function and user-defined function
Introduction to MySQL audit plug-in
phpcms后台上传图片按钮无法点击
【ROS进阶篇】第五讲 ROS中的TF坐标变换
使用 csv 导入的方式在 SAP S/4HANA 里创建 employee 数据
What are the test items of juicer ul982
Qt+pcl Chapter 6 point cloud registration ICP series 4