论文笔记:Parallel Tracking and Verifying: A Framework for Real-Time and High Accuracy Visual Tracking

简介: Parallel Tracking and Verifying: A Framework for Real-Time and High Accuracy Visual Tracking  本文目标在于 tracking performance 和 efficiency 之间达到一种平衡。

Parallel Tracking and Verifying: A Framework for Real-Time and High Accuracy Visual Tracking 

 

本文目标在于 tracking performance 和 efficiency 之间达到一种平衡。将 tracking 过程分解为两个并行但是相互协作的部分:

  一个用于快速的跟踪(fast tracking);

  另一个用于准确的验证(accurate verification)。 

  

本文的 Motivation 主要是:

  1. 大部分跟踪的序列,都是比较平坦简单的,但是存在有些非常具有挑战性的片段的存在,使得跟踪的结果不是非常的好。如果处理不好,还会导致跟踪的丢失。本文利用 verifiers 将进行这些关键点的处理。

  2. 计算机视觉当中多线程计算已经非常普遍,特别是 SLAM。By splitting tracking and mapping into two parallel threads, PTAM (parallel tracking and mapping) [23] provides one of the most popular SLAM frameworks with many important extensions.   

  3. 最近快速、准确的跟踪算法提供了有效的 building blocks,并且鼓励我们去寻找组合的解决方法(呵呵了。。。)

 

 

 

创新点:

  1. we propose to build real-time high accuracy trackers in a novel framework named parallel tracking and verifying (PTAV). 

  2. The key idea is : while T needs to run on every frame, does not. As a general framework, PTAV allows the coordination between the tracker and the verifier: V checks the
tracking results provided by T and sends feedback to V; and V adjusts itself according to the feedback when necessary. By running T and V in parallel, PTAV inherits both the high
efficiency of T and the strong discriminative power of V.

 

==========  分割线  =========

 ======== 以上是 PTAV framework 的流程图,也是两个 tracker 和 verifiers 之间互相协助的过程。 

PTAV Implementation:

1. Tracking 的过程就是利用了 fDSST 跟踪算法,没啥好说的;但是不同的是, the tracker in this paper,存储了所有的中间结果,since sending out last verification request to ensure fast tracing back. 

2. Verifying 是采用了 Siamese network。

  ==>> 当从 tracking 过程中得到的跟踪结果,如果其验证得分低于一个阈值,那么 V 就认为该跟踪结果不可靠,或者说认为已经跟踪失败了。

  此时,V 利用Siamese network,在进行一次检测。具体做法就是利用 region pooling layer 进行一次前传,然后得到许多候选的样本,然后从中选择最好的那个作为检测的结果:

  

  当有了这些检测结果之后,我们在进行一次 check,确认下检测结果是否可信? 其实就是根据检测的置信度和某一阈值进行比较,如果不符合要求,就放大搜索区域,进行再一次的搜索。 

  

 

============================= 算法部分完毕

 实验结果:

 

 想想真可怕,作者居然不辞劳苦的跑了四个数据集。。。

 

相关文章
|
8月前
|
机器学习/深度学习 自然语言处理 算法
Joint Information Extraction with Cross-Task and Cross-Instance High-Order Modeling 论文解读
先前的信息抽取(IE)工作通常独立地预测不同的任务和实例(例如,事件触发词、实体、角色、关系),而忽略了它们的相互作用,导致模型效率低下。
57 0
《Fighting Cybercrime A Joint Task Force of Real-Time Data and Human Analytics》电子版地址
Fighting Cybercrime: A Joint Task Force of Real-Time Data and Human Analytics
57 0
《Fighting Cybercrime A Joint Task Force of Real-Time Data and Human Analytics》电子版地址
|
异构计算
Re12:读论文 Se3 Semantic Self-segmentation for Abstractive Summarization of Long Legal Documents in Low
Re12:读论文 Se3 Semantic Self-segmentation for Abstractive Summarization of Long Legal Documents in Low
Re12:读论文 Se3 Semantic Self-segmentation for Abstractive Summarization of Long Legal Documents in Low
|
机器学习/深度学习 数据采集 人工智能
Re10:读论文 Are we really making much progress? Revisiting, benchmarking, and refining heterogeneous gr
Re10:读论文 Are we really making much progress? Revisiting, benchmarking, and refining heterogeneous gr
Re10:读论文 Are we really making much progress? Revisiting, benchmarking, and refining heterogeneous gr
PAT (Advanced Level) Practice - 1122 Hamiltonian Cycle(25 分)
PAT (Advanced Level) Practice - 1122 Hamiltonian Cycle(25 分)
94 0
PAT (Advanced Level) Practice - 1145 Hashing - Average Search Time(25 分)
PAT (Advanced Level) Practice - 1145 Hashing - Average Search Time(25 分)
100 0
PAT (Advanced Level) Practice - 1118 Birds in Forest(25 分)
PAT (Advanced Level) Practice - 1118 Birds in Forest(25 分)
83 0
PAT (Advanced Level) Practice - 1107 Social Clusters(30 分)
PAT (Advanced Level) Practice - 1107 Social Clusters(30 分)
117 0
PAT (Advanced Level) Practice - 1004 Counting Leaves(30 分)
PAT (Advanced Level) Practice - 1004 Counting Leaves(30 分)
87 0
PAT (Advanced Level) Practice - 1049 Counting Ones(30 分)
PAT (Advanced Level) Practice - 1049 Counting Ones(30 分)
94 0