Abstract

问题：

1)evaluation is often not suffcient

2)biased for certain types of algorthms

3)datasets do not have common ground-truth object positions or extents

4)the initial conditions or parameters of the evaluated tracking algorithms are not the same

本文工作：

1)carry out extensive an evaluation of the state-of-the-art online object-tracking algorithms with various evaluation criteria to understand how these methods perform within the same framework

2)first construct a large dataset with ground-truth objectpositions and extents for tracking and introduce the sequence attributes for the performance analysis

3)second,we integrate most of the publicly available trackers into one code library with uniform input and output formats to facilitate large-scale performance evaluation.

4)Third, we extensively evaluate the performance of 31 algorithms on 100 sequences with different initialization settings.

1 Introduction

定义

难点

总结优劣

数据集

初始化鲁棒性的提出

本文的贡献：基准数据集、代码库、性能评估

2 Brief review of object tracking

1、Representation (描述)

1)holistic(整体) templates

2)LK approaches(do not take large appearance variability into account, not perform well)

3)developed a template update method by exploiting the infrmation of the first frame to correct drifts

4)to better account for appearance changes, subspace-based tracking approaches

5)an efficient LK algorithm and used low-dimensional representations for tracking under varying illumination conditions

6)a robust error norm and proposed an algorithm using a pre-trained view-based eigenbasis representation

7)a low-dimensional subspace representation was learned incrementally to account for target appearance variation for object tracking

8)sparse representations

9)used a dictionary of holistic intensity templates composed of target and trivial templates,and determined the target location by solving multiple l1 minimization problems.

10)local sparse representations and collaborative representations

11) for run-time efficiency, a minimal error bounding strategy was introduced to reduce the number of l1 minimization problems to solve

accelerated proximal gradient approach to efficintly solve l1 minimization problems

12)local sparse appearance model + mean shift algorithm

13)by assuming the representation of particles as jointly sparse ,formulated object tracking as a multi-task sparse learning problem

14)collaborative tracking algorithm=a sparsity-based discriminative classifier+a sparsity-based generative model

15)sparse codes of local image patches with spatial layout in an object were used for modeling the object appearance for tracking

16)a least soft-threshold squares algorithm by modeling image noise with the Gaussian-Laplacian distribution other than the trivial templates

17)a number of tracking methods based on color histograms

18)Recently, discriminative models have been developed in the field of object tracking,where a binary classifier is learned online to separete the target from the background.Numerous classifier object tracking.

19)To account for an appearance change caused by a large pose variation and heavy occlusion, an object can be represented by parts with descriptors or histograms.

20)Several approaches based on multiple representation schemes have been developed, to better handle appearance varations.

2、 Search Mechanism

deterministic and stochastic search methods have been developed to estimate the object states.

...

objective functions for object tracking are usually nonlinear with many local minma.To alleviate this problem, dense sampling metnods have been adopted,at the expense of a high computational load.Onthe other hand, stochastic search algorithms such as particle filters have been widely used since they are relatively insensitive to the local minimum and are computationally efficeint.

3、Model Update

1)online update of target representation to account for appearance variations plays an important role for robust object tracking.

2)...addressed the template update problem for the LK algorithm,where the template was updated with the combination of the fixed reference template extracted from the first frame and the result from the most recent frame.

3)effective update algorithms have also been proposed in the from of the online mixture model, online boosting, and incremental subspace update.

4)discriminative model:recently, considerable attention has been paid to draw samples effective for training online classifiers.

5)semi-supervised

6)someone focused on the tracking problem within the multiple instance learning framework and developed an online algorithm.

7)to exploit the underlying structure of the unlabeled data, Kalal et al. developed a tracking algorithm within the semi-supervised learning framework to select positive and negative samples for model update.

8)the proposed tracking algorthm directly predicts the target location change between frames on the basis of structured learning.

9)a tracking method based on co-training to combine generative and discriminative models.

4、 Context and Fusion of Trackers

5、Performance Evaluation

1)time-reversed Markov chain

2)introduced a unified conceptual framework and presented an experimental analysis.

3)poor initialization of a tracker signigicantly decreases the tracking accuracy, however, further analysis based on comprehensive experimental evaluations is necessary and important to better understand the state-of-the-art algorithms.

4)a ranking approach to analyze the reported results of object traching methods.

5)the failure rate of a tracking method was computed by counting the number of frames in which a method fails to follow a target object.

6、Challenging Factors

1)Occlusion

2)Deformation

　　　　modeled the target appearance by using a small number of rectangular blocks from which histograms were extracted. The positions of these blocks within an object were adaptively determined for object tracking

　　　　a target object was represented by a patch-based appearance model and the topology between local patches was updated online.

　　　　based on segmentation techniques to describe opject shape.

　　　　based on a generalized Hough transform and used segmentation based on the GrabCut method to better describe the forground objects.

3)Scale Variation

　　　　search at multiple scales and use the one with the maximum likelihood for tracking 

　　　　used the scale space theory to improve the mean-shift tracking method

　　　　include object scale as one state in the motion model 

　　　　in the tracking methods based on particle filters, object states are often estimated by the average of a few particles with large weights

4)Fast Motion

　　　　extended the mean-shift tracking method by using multiple kernels centered around fast motion areas

　　　　introduced the Wang-Landau Monte Carlo sampling method to handle fast motion by alleviating motion smoothness constraints with both the likelihood functions and the desity of states.

　　　　to cope with abrupt motion and large appearance changes, multiple trackers with different motion and appearance models were used where the best one was selected using Markov Chain Monte Carlo sampling.

3 Evaluated tracking algorithms

as all implementations inevitably involve technical details and specific parameter settings, we included the algorithms only if the original source or binary code was publicly available

4 datasets

VIVID

CAVIAR

TB-100

TB-50(challenging)

attributes of a test sequence(11个)

5 Evaluation methodology

position accuracy robustness over a certain type of appearance changes, tracking speed, memory requirement, and ease of use can be considered.

precision plot:中心距离，尺寸如何体现：阀值

success plot：AOS,AUC(召回率？)

1)Robustness evaluation:OPE(one-pass evaluation)->TRE(temporal robustness evaluation) & SRE(spatial robustness evaluation)

2)Robustness Evaluation with Restart（下面是承接关系，不是并列关系）:

　　OPER(One-Pass Evaluation with Restart)

　　SRER(Spatial Robustness Evaluation with Restart)

　　　　as in the case of OPER, we evaluate whether a tracking method is sensitive to spatial perturbatio with restarts such that the tracking  　　　　performance in challenging sequences can be better analyzed.

　　Approximation using Virtual Runs

　　　　两个问题：1、restart的阀值在不同情况下选择不同，选所有情况的阀值不切实际；2、很多算法提供binary code，无法检测到失败然后重启。

　　　　因此，提出该法：...（论文中有一个说明例子）

6 Evaluation Results

1)overall performance

2)performance of SRER

3)performance Analysis by Attributes

4)tracking speed

7 Conclusions

参考文献：Yi Wu, Jongwoo Lim, and Ming-Hsuan Yang. Object Tracking Benchmark.***