论文笔记 - ATOM: Accurate Tracking by Overlap Maximiza

文件名称: 论文笔记 - ATOM: Accurate Tracking by Overlap Maximization.pdf

所属分类: 讲义

开发工具:

文件大小: 1mb

下载次数: 0

上传时间: 2019-10-13

提供者: weixin_********

下载 (1mb)

不能下载？报告错误

详细说明：论文笔记 - ATOM: Accurate Tracking by Overlap Maximization2019/10/13 机器人针对服务机器人系统对环境感知问题,对人和其他目标的状态预测视频监控 2819 obiect 农险管慧云再 2019/10/13 虚拟现实人机交互 7 视频剪辑加班加点扣掉吴秀波总算是快弄完了,P了三天三夜应该是不仔细找都看不出来了真的惨,怎么刚入职就碰上这种事唉,我尽力了 P了二天三夜 2019年1月31日01:46 如果楼上学会目國■以标跟踪 2019/10/13 Bilibili智能防挡弹幕 00:05■结束原你坤/全民制作人们大冢好一个人生不尽兴?传送门https://www.bilibili.com/video/av61335515?from=search&seid=18435316748105582121 9 02 难点 ◇ 令 10 2019/10/13 尺度 KCF Tracker FPS c2 11 模糊 KCF Tracker FPS: 388 12 2019/10/13 姿态 A ROI sector 消失 KCF Tracker FPS: 108 14 2019/10/13 遮挡 KCF Rocket 15 03 算法 ◇ 令 16 2019/10/13 目标跟踪图谱:https://github.com/horwood/benchmarkresults./tree/master denotes VOT baseline experiment ccor為20102017 denotes VoT rea/lime experirmerIl SRDCE2015园4 FGF况07E DeOr SRDCF A: 2015 2 ATOM SPD-Fgcor CsRECE多2:7 OPR RI-MDNet DAT MDNetA 20151 cNNl-svI KCF 20143 MCPF CN加158 ROLO CSK K二FMr GOTUJRN LE MOSS SiamAk DSST DaSiatnRPN乡 StelInRPN Learno? SAMF 2 BAcE Siamese Cende RPN MCCE DsNet OTB2013 OT82015 Srmeu ciompoerc43c172 EAST VOT wa015 wo2016 -2017 LSYC D → ube boundiegBoHD rclated datasct Benchmark Temrlecolrr126 NUS P Tracking: L234V LTW: LnDlerm acorn n the wHl ATOM: Accurate Tracking by Overlap Maximization 论文链接:htps:/ xiv.org/pdt/181107628pdt 源码链接:htts:/github.com/visionmlpytracking 论文作者: Martin danelljan, Goutam bhat(cVL, Linkoping∪ nIversity, Sweden)等人论文出处:CvPR2019 该文章再一次将单目标跟踪的精确度拔高到了一个新的高度。本文的主要思想是通过 oU-Net来优化网络预测的 bounding box,提高定位精确度;并通过基于共轭梯度的策略来训练分类网络,提高分类准确率。通过二者的结合,ATOM在四个主要的跟踪数据集上达到了新的高度。 18 2019/10/13 摘要 While recent years have witnessed astonishing improvements in在近些年,单目标跟踪算法大多集中在开发 visual tracking robustness, the advancements in tracking accuracy 比较鲁棒的分类器方面,在 bounding boxl的 have been limited. As focus has been directed towards the refine没有太多进展,大多是通过muti-Sca|e development of powerful classifiers the prob1mn, ccurate target state的输入,来找到一个响应值最高的输入,作 overlooked. In fact, most trackers reot, to a simple multi- scale search为 ounding box的 scale。但是,单目标跟踪 in order to estimate the target bounding box. We argue that this 的物体有些是非刚体,随着物体姿态的变化 approach is fundamentally limited since target estimation is a comp!lr物休会发生长宽比的改变,因此,单纯计算 task, requiring high-level knowledge about the object sca|e并不能得到一个很好的 bounding box. 19 对比 comparison of our approach with state-of-the-art trackers UPDT, based on correlation filters, lacks an explicit target state estimation component, performing a brute-force multi cale search instead. C ntly, it does not handle aspect-ratio changes, which can lead to tracking failure (second row). DaSiamRPN employs a bounding box regression strategy to estimate the target state, but still struggles in cases of out-of-plane rotation, deformation, etc Our approach ATOM, employing an overlap prediction network, successfully handles these challenges and provides ATOM DasiamRPl accurate bounding box predictions

(系统自动生成,下载前可以参看下载内容)