2012-ECCV-Low-rank sparse learning for robust visual tracking

Zhang2012LowRankSL.md

2012 - ECCV - Low-rank sparse learning for robust visual tracking

Abstract

• propose a new particle-filter based tracking algorithm that exploits the relationship between particles (candidate targets)
  • 就是 low-rank

Introduction

• visual tracking exploits the sparse representation of the target candidate using a dictionary of templates
  • advantage
    • robustness to a wide range of image corruptions, especially moderate occlusions
  • disadvantage
    • a computational cost dominated by $\ell_1$ minimization, and the computational cost grows linearly with the number of particle samples used.
    • learns the sparse representations of particles separately and thus ignores the relationships among them
      • 这一点是本文的 Motivation
• 作者这里说的 relationship among them 指的什么？
  • most particles are densely sampled at a small distance around the target, so their appearances, and thus their sparse representations, are expected to be similar.
• 本文方法
  • adopt the independent and identically distributed (i.i.d.) sampling strategy
  • The next object state is chosen to be the particle that has the highest similarity with a dictionary of target object templates.
    • 因此 this dictionary is crucial
• 本文方法 based on 的 observations
  • The best particle sample should have the most similar representation with the target object templates in the dictionary.
    • 这点其实在说 稀疏性吧
  • Since particles are densely sampled around the current target state, the appearances of many of these particles and, in turn, their representations w.r.t to the dictionary are expected to be similar. So, we observe a correlation among particle representations and not their i.i.d. sampled states.
    • 这点在说 correlation matrix 的 low-rank 性质
  • To alleviate the effect of Occlusion and noise, representation error should be explicitly incorporated in the tracking process.
    • 这点是在说 template update
  • During tracking, a particle sample should be represented using a dictionary of templates composed of both object and background templates, which are updated progressively.
    • This emphasizes the importance of representing what a target is and is not.
    • Discriminating the target from the background adds another layer of robustness against possible tracker drift.
    • dictionary 既要包含 object templates 也要包含 background templates
    • 这是 2009 - ICCV - Robust visual tracking using L1 minimization 里面没有的，2009 - ICCV - Robust visual tracking using L1 minimization 只是提到了用比较中立的 trivial templates

Related Work

• Generative tracking methods
  • adopt an appearance model to describe the target observations, and the aim is to search for the target location that has the most similar appearance to the model
    • 和给定的 model 最相似，重点在 similarity measure
  • eigentracker [6]
    • IJCV - 1998 - Eigentracking: Robust matching and tracking of articulated objects using a view-based representation
  • mean shift tracker [7]
    • TPAMI - 2005 - Kernel-Based Object Tracking
  • incremental tracker [8]
    • IJCV - 2008 - Incremental Learning for Robust Visual
• Discriminative tracking methods
• view the object tracking as a binary classification problem, which seeks the target location that can best separate the target from the background.
  • 最能分离 target 和 background，重点在 discrimination
• on-line boosting [9]
  • BMVC - 2006 - Real-Time Tracking via On-line Boosting
• ensemble tracking [10]
  • CVPR - 2005 - Ensemble tracking
• online multi-view forests for tracking [11]
  • LNCS - 2010 - On-Line Multi-view Forests for Tracking
• online multiple instance learning tracking [12]
  • CVPR - 2009 - Visual tracking with online multiple instance learning

Low-Rank Sparse Tracker (LRST)

Adaptive Dictionary: $D_t$

• the tracking result $\textbf{y}t$ at instance $t$ is the particle $x_i$ such that $i = \arg \max{k=1,\cdots,n} (\Vert z_k^O\Vert_1 - \Vert z_k^B\Vert_1)$
• This encourages the tracking result to be represented well by the object and not the background templates.
• exploit this discriminative information to design a systematic procedure for updating $D_t$
  • To each object template in $D_O$, we allocate a weight $ω_i$ that is indicative of how representative the template is. In fact, the more a template is used to represent tracking results, the higher its weight is.
  • If particles are suffciently represented (up to a predefined threshold) by the dictionary, then there is no need to update it. Otherwise, the current tracking result replaces the object template that has the smallest weight.

@inproceedings{Zhang2012LowRankSL,
  title={Low-Rank Sparse Learning for Robust Visual Tracking},
  author={Tianzhu Zhang and Bernard Ghanem and Si Liu and Narendra Ahuja},
  booktitle={ECCV},
  year={2012}
}