[论文阅读 2019 综述目标跟踪]Deep Learning for Visual Tracking: A Comprehensive Survey

最新推荐文章于 2022-07-08 11:16:58 发布

lingqing97

最新推荐文章于 2022-07-08 11:16:58 发布

阅读量404

点赞数 1

分类专栏：论文阅读文章标签：深度学习人工智能

本文链接：https://blog.csdn.net/qq_39621037/article/details/114712428

版权

19 篇文章 5 订阅

订阅专栏

这篇论文是之前看的，主要是对深度学习下跟踪算法的一个综述。

由于是综述，内容涵盖较多，这里截取部分我感兴趣的内容。

首先，作者根据多个方面对跟踪算法进行了一个分类，如下图所示:

在这里插入图片描述

然后，论文中还介绍了一些在单目标跟踪中常用的评价指标:

Center Location error(CLE): The CLE or precision metric is defined as the average Euclidean distance between the precise ground-truth locations of the target and estimated locations by a visual tracker(简单说就是预测位置与标签的欧式距离误差)
Accuracy: The accuracy is then calculated by the average overlap scores (AOS) during the tracking when a visual tracker’s estimations have overlap with the ground-truth ones(即average IOU)
Robustness/failure score:The robustness or failure score is defined as the number of required reinitializations when a tracker loses (or drifts) the target during the tracking task. The failure is detected when the overlap score drops to zero.(Robustness是跟踪器跟踪失败后重新初始化的次数，failure是失败总数，当IOU score为0时判定为failure)
Expected average overlap(EAO):the EAO score is calculated as $\widehat{\Phi}_{N_{s}}=\left\langle\frac{1}{N_{s}} \sum_{i=1}^{N_{s}} \Phi_{i}\right\rangle$ ,where $\Phi_{i}$ is defined as the average of perframe overlaps until the end of sequences, even if failure leads to zero overlaps.