supplementary ﬁle for robust visual tracking using oblique ......for the cases with a large number...

Supplementary file for Robust Visual Tracking Using Oblique Random Forests

Le ZhangADSC

Singapore [email protected]

Jagannadan VaradarajanADSC


Ponnuthurai Nagaratnam SuganthanNTU


Pierre MoulinUIUC

IL 61820-5711 [email protected]

Narendra AhujaUIUC

IL 61820-5711 [email protected]

In this document we first provide the derivation for incremental learning of PSVM parameters, which is used to updateevery decision node in our online Obli-RaF. Next, we provide a detailed analysis of various parameters of the proposedtracker, and demonstrate the merits of our Obli-RaF. Finally, we provide more detailed results of our approach on the OTB-51, OTB-100 datasets and VOT2016 under various challenging scenarios.

1. Online update of PSVM parametersLet X = [x1, . . . ,xN ]> ∈ IRN×M be an N ×M matrix obtained by stacking N samples in X . The non-italicised >

indicates a transpose operation. Let Y ∈ {−1,+1}N be a vector obtained by stacking the labels of samples in X . Here, wedrop the time index t in Xt, since the following formulation applies to any t. We define a diagonal matrix D, whose diagonalentries Di,i = 1 if xi belongs to the positive class and -1 otherwise. The parameters of PSVM are obtained by solving thefollowing problem:

min(w,b,ξ)

1

2||ξ||2 + ν ∗ 1

2(w>w + b2)

s.t.D(Xw − be) + ξ = e.

(1)

where ξ is the error vector and ν is the regularization parameter, w and b are the coefficient of the hyperplane, and e is avector of all ones. The parameters of PSVM {w, b}, can be computed in closed form using the following formulation:

[w; b]> = (νI + H>H)−1H>De; H = [X,−e] (2)

Let De = Y and suppose at time t, we have the solution βt = [wt, bt]>. The parameters βt+1 corresponding to time

t + 1 can be calculated using βt and newly available data without directly solving Eq. (2). The key step in updating theparameters is the calculation of (νI + H>H)−1. Suppose the features corresponding to data at time t+ 1 is Ht+1, then theproblem becomes

minimizeβt+1

∥∥∥∥[ Ht

Ht+1

]βt+1 −

[Yt

Yt+1

]∥∥∥∥2

+ ν||βt+1||2, (3)

It is straightforward to observe that

βt+1 =

([Ht

Ht+1

]> [Ht

Ht+1

]+ νI

)−1 [Ht

Ht+1

]> [Yt

Yt+1

]

= S−1t+1

[Ht

Ht+1

]> [Yt

Yt+1

],

(4)

1

where

St+1 =

[Ht

Ht+1

]> [Ht

Ht+1

]+ νI, (5)

Then,

St+1 =

[Ht

Ht+1

]> [Ht

Ht+1

]+ νI = St + H>t+1Ht+1, (6)

And, [Ht

Ht+1

]> [Yt

Yt+1

]= H>t Yt + H>t+1Yt+1,

= StS−1t H>t Yt + H>t+1Yt+1,

= Stβt + H>t+1Yt+1,

= (St+1 −H>t+1Ht+1)βt + H>t+1Yt+1,

= St+1βt −H>t+1Ht+1βt + H>t+1Yt+1.

(7)

From Eqn. (4),

βt+1 = S−1t+1

[Ht

Ht+1

]> [Yt

Yt+1

]= S−1

t+1(St+1βt −H>t+1Ht+1βt + H>t+1Yt+1)

= βt + S−1t+1H

>t+1(Yt+1 −Ht+1βt).

(8)

From matrix inversion lemma, we have

S−1t+1 = (St + H>t+1Ht+1)−1

= S−1t − S−1

t H>t+1(I + Ht+1T−1t H>t+1)−1Ht+1S

−1t ,

(9)

To summarize, let ϕt = S−1t , the on-line update of the weight can be achieved by the following operation:

ϕt+1 = ϕt −ϕtH>t+1(I + Ht+1ϕtH

>t+1)−1Ht+1ϕt

βt+1 = βt +ϕt+1H>t+1(Yt+1 −Ht+1βt).

(10)

2. Complexity AnalysisSuppose we have N data samples in total and at each node Ms � M features are selected to search for the decision

hyperplane. Without loss of generality, we assume each tree grows to its largest extent with a maximum depth Γdepth, wherethe root node is at level 0 and the leaf nodes is at level Γdepth. In order to simplify the analysis, we further assume that eachtree is balanced, i.e., each decision node will divide the data into two branches with equal size. In this case, 2Γdepth = N .

In a binary tree, there are 2i nodes at any depth i with each node associated with N2i samples. The orthogonal split

algorithm firstly ranks each feature and finds a threshold to optimize the impurity criterion. Regardless of the computationaltime for calculating the impurity value, the computation complexity to find a “best-split” in an orthogonal decision tree takestime on the order of Ms

N2i log N

2i , where N2i log N

2i is the cost of sorting each feature. From this, it is straightforward to showthat the total complexity of a (balanced) orthogonal decision tree is:

COrth =

L∑i=0

2i(MsN

2ilog

N

2i); (11)

The overall complexity of an orthogonal decision tree, which is denoted by COrth, is around O(MsN log2N). With com-plexity of PSVM being of order M3

s , the overall complexity of an oblique decision tree is given by:

CObli = (2N − 1) ∗M3s , (12)

which is about O(N ∗M3s ). But typically, only a few features are sampled at each internal node, resulting in Ms � logN

for the cases with a large number of training samples (such as visual tracking where data samples accumulate). Thus, we getCObli � COrth.

2.1. Obli-RaF improves ConNets

Our Obli-RaF+ConvNet tracker can be regarded as a combination of the simple Obli-RaF tracker and FCNT tracker [1],as mentioned in the main paper. These two models work collaboratively to improve each other. This allows high performanceto be achieved since it becomes unlikely for compatible classifiers trained on independent features to agree on an incorrectlabel. Table 1 support our hypothesis by providing detailed results on the precision value of our method and the FCNT modelwhen the threshold for the error is set to be 20 pixels, as suggested in [2].

Dataset Ours FCNTcarDark 100.00 100.00

car4 100.00 100.00david 99.79 93.84david2 100.00 100.00

sylvester 98.44 99.26trellis 99.82 98.59fish 100.00 100.00

mhyang 99.93 99.93soccer 52.81 43.62matrix 80.00 2.00

ironman 64.46 6.63deer 98.59 100.00

skating1 100.00 65.50shaking 98.36 92.05singer1 99.15 92.88singer2 87.98 80.87

coke 89.35 90.03bolt 87.71 96.57boy 100.00 100.00

dudek 93.45 91.18crossing 100.00 100.00couple 100.00 100.00

football1 100.00 56.76jogging-1 97.39 97.39jogging-2 96.09 95.44

Dataset Ours FCNTdoll 96.82 98.68girl 99.80 91.20

walking2 100.00 100.00walking 100.00 100.00fleetface 74.68 75.53freeman1 100.00 100.00freeman3 100.00 91.74freeman4 99.65 99.65

david3 94.05 93.65jumping 100.00 99.68carScale 82.54 73.81skiing 100.00 100.00dog1 100.00 100.00suv 92.91 93.54

motorRolling 92.07 90.85mountainBike 99.56 100.00

lemming 85.85 43.26liquor 68.58 87.02

woman 91.96 91.96FaceOcc1 53.14 54.15FaceOcc2 84.36 85.47basketball 86.21 85.79football 96.96 96.69subway 100.00 99.43tiger1 71.92 71.92

Dataset Ours FCNTtiger2 73.97 80.27biker 50.70 50.70bird1 36.52 34.31bird2 92.93 94.95

blurBody 70.36 63.17blurCar1 94.47 93.13blurCar2 86.32 92.82blurCar3 98.04 98.88blurCar4 92.63 86.58blurFace 99.59 99.59blurOwl 87.64 94.77

board 85.67 31.95bolt2 89.42 96.25box 78.90 88.46car1 76.47 18.43car2 100.00 100.00car24 99.84 93.69

clifBar 75.85 56.14coupon 100.00 39.45crowds 100.00 12.68dancer 98.67 98.67

dancer2 100.00 99.33diving 11.63 12.09

dog 100.00 98.43dragonBaby 92.04 86.73

Dataset Ours FCNTgirl2 9.33 7.20gym 72.88 94.78

Human2 64.18 65.69Human3 79.27 98.00Human4 50.67 81.56Human5 51.47 49.37Human6 29.17 30.68Human7 98.80 78.40Human8 98.44 89.84Human9 83.28 85.90

jump 8.20 9.84kiteSurf 100.00 91.67

man 100.00 100.00panda 100.00 100.00

redTeam 100.00 100.00rubik 89.98 89.63skater 98.75 95.00skater2 94.48 82.53

skating2-1 25.16 27.70skating2-2 36.36 34.46

surfer 99.73 100.00toy 97.79 93.73

trans 38.71 47.58twinnings 100.00 100.00

vase 79.70 81.18

Table 1. Effect of combining Obli-RaF with FCNT.Each entry in the table stands for the mean avrage precision on each dataset.

3. Sensitivity analysisThe free parameters involved in our model are: i) depth of the tree Γdepth, ii) maximum number of trees in the forest K,

iii) number of features sampled for a node Ms, and iv) regularization parameter ν. We demonstrate the result with the simpleObli-RaF tracker. Similiar conclusion can be applied to the ConvNets based Obli-RaF tracker.Tree depth. We find that in general Obli-RaF results in shallow decision trees. Therefore, setting Γdepth to 3 is mostlysufficient. However, we set Γdepth = 100 which does not impact the performance.Number of features. The number of features Ms used in each node controls the diversity of each node in Obli-RaF. A smallvalue for Ms results in less-correlated decision trees whereas large values reduce the diversity of the tree ensembles as mostof the trees tend to agree on unseen test data. From our experiments, setting Ms =

√M gives the best performance.

Number of trees. Typically, increase in the ensemble size K provides more generalization ability of Obli-RaF. However,setting K to be too large increases the computational complexity. We experimented with a range of values for K as shown inFig. 1 and found that K = 100 gives satisfactory results.Effect of ν. The regularization parameter ν in Eqn.1 allows us to adjust the trade-off between complexity and trainingaccuracy of PSVM. While it may be beneficial to tune this parameter for each node, we simply set it to 100. Generallyspeaking, a large value of ν will favor a hyperplane with less complexity while a small value results in over fitting. Weexperimented by varying ν between 10 and 1000 and the results are summarized in Fig. 2.

4. Advantages of Oblique Random forest

Obli-RaF learns shallow trees. Our tree induction method results in shallow trees and thus, more efficient. On average, the

0 10 20 30 40 500

0.2

0.4

0.6

0.8

1

Location error threshold

Pre

cis

ion

Precision plots of OPE

Obli−RaF_150 [0.803]

Obli−RaF [0.800]

Obli−RaF_50 [0.757]

Obli−RaF_80 [0.755]

0 0.2 0.4 0.6 0.8 10

0.2

0.4

0.6

0.8

1

Overlap threshold

Success r

ate

Success plots of OPE

Obli−RaF_150 [0.582]

Obli−RaF [0.580]

Obli−RaF_50 [0.552]

Obli−RaF_80 [0.547]

Figure 1. Average success plot and precision plot of Obli-RaF tracker with different values for K. Obli-RaF indicates setting K = 100.Obli-RaF 50, Obli-RaF 80 and Obli-RaF 150 stand for K = 50, 80 and 150 respectively.

0 10 20 30 40 500

0.2

0.4

0.6

0.8

1

Location error threshold

Pre

cis

ion

Precision plots of OPE

Obli−RaF[0.800]

Obli−RaF_nu_1000[0.784]

Obli−RaF_nu_10[0.767]

0 0.2 0.4 0.6 0.8 10

0.2

0.4

0.6

0.8

1

Overlap threshold

Success r

ate

Success plots of OPE

Obli−RaF[0.580]

Obli−RaF_nu_1000[0.564]

Obli−RaF_nu_10[0.555]

Figure 2. Average success plot and precision plot of RaF tracker with different ν parameter. Obli-RaF indicates setting ν = 100. Obli-RaF nu 10 and Obli-RaF nu 1000 denotes ν = 10 and 1000 respectively.

Figure 3. Visualization of different decision trees. (left) proposed oblique decision tree; (right) conventional decision tree. Leaf nodes aredenoted in blue. The first and second number in the bracket denote the number of samples from the first and second class within that noderespectively. Oblique decision trees result in shallow trees.

proposed simple Obli-RaF tracker runs 3 times faster than the orthogonal one. In order to show this, we crop 500 positivetraining samples and 500 negative samples from the first frame of the “car4” video sequence and train an oblique and anorthogonal decision tree. The learned trees are shown in Fig. 3. As we can see, the oblique decision tree (on the left)

classifies the samples correctly with a tree depth of three by using only two clustering operations, whereas the orthogonaldecision tree (on the right) needs 5 levels and several orthogonal splits to achieve the same result.

Fig. 3 shows histograms of the tree depth of oblique and orthogonal RaF when applied on the OTB50 benchmark videos.From this, we clearly see that the average depth of an oblique trees is 1.51 whereas the average depth of orthogonal decisiontrees is 8.86. The maximum tree depth obtained for oblique and orthogonal decision trees are 7.16 and 22.86 respectively.We also calculated the ratio of tree depth obtained from orthogonal and oblique RaF for each frame in the benchmark andobtained an average depth ratio of 5.89, which denotes that on average the oblique decision tree is around 5.89 times shallowerthan the orthogonal one. From these results, we can conclude that our PSVM based Obli-RaF can better capture the geometricstructure of the data samples and yield shallower trees.

0 5 10 15 20 250

1000

2000

3000

4000

5000

6000

7000

8000

Oblique RaF

Orthogonal RaF

Figure 4. The histogram of the average tree depth of the forest.

Obli-RaF learns smoother decision boundaries. Here, we choose a toy example to demonstrate the the advantagesof our method compared to conventional random forest. To this end, a three class artificial spirals dataset was generatedas shown in the first part Fig 5(a). Fig. 5(b) demonstrates the hierarchical partitions from a single axis-parallel decisiontree. We can clearly see that the spiral decision boundary is approximated by many “stage-like” decision hyperplanes.The decision boundary can be somehow smoothed by averaging many decision trees, which is demonstrated in Fig 5(c).The decision boundary in this case is obtained by a conventional orthogonal random forest (Orth-RaF). Fig 5(d) shows thedecision boundary from the proposed Obli-RaF, which fits the data much better than the former two methods resulting in a“smoother” decision boundary.

−10 −5 0 5 10

−10

−5

0

5

10

A Toy Example

x1

x2

Class 1

Class 2

Class 3

(a) a

−10 −5 0 5 10

−10

−5

0

5

10

x1

x2

Decision Boundary of a Single DT

(b) b

−10 −5 0 5 10

−10

−5

0

5

10

x1

x2

Decision Boundary of Orth−RaF

(c) c

−10 −5 0 5 10

−10

−5

0

5

10

x1

x2

Decision Boundary of Obli−RaF

(d) d

Figure 5. Decision boundary of the spirals dataset learned by different methods. (a) Ground truth, decision boundary from a (b) singledecision tree, (c) Orthogonal RaF, (d) Oblique RaF.

5. Detailed resultsIn this section we provide detailed results of from our collaborative tracker (Obli-RaF tracker with ConvNets) under each

challenging scenarios within OTB-51 and OTB-100 datasets. We present the OPE, SRE and TRE results for both OTB-51 andOTB-100 in Fig. 6, 7, 8, 9, 10, 11 respectively. The precision curves indicate that our method works remarkably well undermany challenging scenarios. For instance, our method outperforms state-of-the-art methods under illumination variations,out-of-plane rotation, out-of-view, background cluttering and low resolution.

We also evaluate the proposed hybrid tracker on VOT2016 which consists of 62 datasets and 74 trackers . The AR rankingof all the methods are summarized in Table 2. Note we adopt the same parameter setting as done in OTB without any furthertuning.

Location error threshold0 10 20 30 40 50

Pre

cis

ion

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1Precision plots of OPE

Ours [0.919]

HCF [0.891]

MUSTer [0.865]

FCNT [0.864]

LCT [0.848]

SRDCF [0.838]

MEEM [0.830]

KCF [0.741]

DSST [0.739]

TGPR [0.705]


Pre

cis

ion

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1illumination variation (23)

Ours [0.881]

HCF [0.831]

FCNT [0.786]

MUSTer [0.777]

LCT [0.774]

MEEM [0.750]

SRDCF [0.740]

DSST [0.718]

KCF [0.706]

TGPR [0.644]


Pre

cis

ion

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1out-of-plane rotation (37)

Ours [0.908]

HCF [0.863]

MUSTer [0.844]

LCT [0.843]

MEEM [0.833]

FCNT [0.831]

SRDCF [0.809]

DSST [0.761]

KCF [0.759]

TGPR [0.682]


Pre

cis

ion

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1scale variation (28)

Ours [0.921]

HCF [0.880]

FCNT [0.832]

MUSTer [0.817]

MEEM [0.785]

SRDCF [0.778]

LCT [0.758]

DSST [0.740]

KCF [0.680]

SCM [0.672]


Pre

cis

ion

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1occlusion (27)

HCF [0.869]

Ours [0.869]

MUSTer [0.845]

LCT [0.836]

SRDCF [0.833]

KCF [0.790]

FCNT [0.788]

MEEM [0.786]

DSST [0.763]

TGPR [0.680]


Pre

cis

ion

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1deformation (17)

Ours [0.919]

FCNT [0.898]

HCF [0.868]

LCT [0.861]

MUSTer [0.845]

SRDCF [0.840]

MEEM [0.832]

KCF [0.804]

DSST [0.710]

TGPR [0.700]


Pre

cis

ion

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1 motion blur (12)

HCF [0.844]

Ours [0.824]

SRDCF [0.789]

FCNT [0.784]

MEEM [0.715]

MUSTer [0.695]

LCT [0.664]

KCF [0.650]

DSST [0.603]

Struck [0.551]


Pre

cis

ion

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1fast motion (17)

Ours [0.835]

HCF [0.790]

MEEM [0.742]

SRDCF [0.741]

FCNT [0.734]

MUSTer [0.695]

LCT [0.665]

Struck [0.604]

KCF [0.602]

DSST [0.562]


Pre

cis

ion

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1in-plane rotation (31)

Ours [0.908]

HCF [0.868]

FCNT [0.839]

LCT [0.802]

MEEM [0.800]

MUSTer [0.799]

DSST [0.780]

SRDCF [0.766]

KCF [0.725]

TGPR [0.675]


Pre

cis

ion

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1Precision plots of OPE - out of view (6)

Ours [0.799]

LCT [0.728]

MEEM [0.727]

MUSTer [0.709]

HCF [0.695]

SRDCF [0.680]

FCNT [0.670]

KCF [0.649]

TLD [0.576]

Struck [0.539]


Pre

cis

ion

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1background clutter (21)

Ours [0.911]

HCF [0.885]

MUSTer [0.831]

FCNT [0.805]

SRDCF [0.803]

MEEM [0.797]

LCT [0.796]

KCF [0.752]

TGPR [0.717]

DSST [0.691]


Pre

cis

ion

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1Plow resolution (4)

Ours [0.999]

FCNT [0.978]

HCF [0.935]

MEEM [0.888]

SRDCF [0.767]

MUSTer [0.750]

DSST [0.738]

LCT [0.717]

SCM [0.661]

KCF [0.630]

Overlap threshold0 0.2 0.4 0.6 0.8 1

Success r

ate

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1Success plots of OPE

MUSTer [0.641]

LCT [0.628]

SRDCF [0.626]

Ours [0.615]

FCNT [0.606]

HCF [0.605]

MEEM [0.566]

KCF [0.513]

DSST [0.505]

TGPR [0.503]


Success r

ate

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9


Ours [0.599]

MUSTer [0.595]

LCT [0.575]

FCNT [0.563]

HCF [0.560]

SRDCF [0.556]

MEEM [0.532]

DSST [0.502]

KCF [0.490]

TGPR [0.476]


Success r

ate

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9


LCT [0.617]

MUSTer [0.609]

Ours [0.606]

SRDCF [0.591]

FCNT [0.580]

HCF [0.579]

MEEM [0.553]

KCF [0.514]

DSST [0.508]

TGPR [0.486]


Success r

ate

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9


Ours [0.598]

MUSTer [0.594]

SRDCF [0.587]

FCNT [0.556]

LCT [0.553]

HCF [0.531]

SCM [0.518]

MEEM [0.498]

DLT [0.458]

DSST [0.451]


Success r

ate

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1occlusion (27)

MUSTer [0.623]

SRDCF [0.618]

LCT [0.618]

HCF [0.596]

Ours [0.588]

FCNT [0.565]

MEEM [0.545]

KCF [0.540]

DSST [0.504]

TGPR [0.485]


Success r

ate

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1deformation (17)

LCT [0.659]

MUSTer [0.641]

FCNT [0.640]

Ours [0.623]

SRDCF [0.622]

HCF [0.612]

KCF [0.577]

MEEM [0.551]

TGPR [0.515]

DSST [0.511]


Success r

ate

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1motion blur (12)

HCF [0.616]

Ours [0.607]

SRDCF [0.601]

FCNT [0.589]

MUSTer [0.558]

MEEM [0.541]

LCT [0.524]

KCF [0.499]

DSST [0.458]

TGPR [0.434]


Success r

ate

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1fast motion (17)

Ours [0.604]

HCF [0.578]

SRDCF [0.569]

MEEM [0.553]

MUSTer [0.552]

FCNT [0.547]

LCT [0.534]

Struck [0.462]

KCF [0.461]

DSST [0.433]


Success r

ate

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9


Ours [0.600]

LCT [0.592]

MUSTer [0.582]

HCF [0.582]

FCNT [0.576]

SRDCF [0.566]

MEEM [0.535]

DSST [0.532]

KCF [0.497]

TGPR [0.479]


Success r

ate

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1out of view (6)

MEEM [0.606]

Ours [0.596]

LCT [0.594]

MUSTer [0.590]

HCF [0.575]

SRDCF [0.555]

KCF [0.550]

FCNT [0.519]

DSST [0.490]

Struck [0.459]


Success r

ate

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9


HCF [0.623]

MUSTer [0.617]

Ours [0.599]

SRDCF [0.587]

LCT [0.587]

MEEM [0.569]

FCNT [0.568]

KCF [0.533]

TGPR [0.522]

DSST [0.492]


Success r

ate

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1low resolution (4)

Ours [0.491]

SRDCF [0.471]

FCNT [0.466]

HCF [0.442]

SCM [0.438]

MEEM [0.401]

MUSTer [0.396]

LCT [0.386]

DSST [0.352]

DLT [0.330]

Figure 6. Detailed Results of OPE on OTB-51 benchmark.

References[1] L. Wang, W. Ouyang, X. Wang, and H. Lu. Visual tracking with fully convolutional networks. In IEEE International Conference on

Computer Vision, pages 3119–3127, 2015. 3[2] Y. Wu, J. Lim, and M.-H. Yang. Online object tracking: A benchmark. In Proceedings of the IEEE Conference on Computer Vision

and Pattern Recognition, pages 2411–2418, 2013. 3


Pre

cis

ion

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1Precision plots of TRE

HCF [0.871]

Ours [0.865]

FCNT [0.846]

MEEM [0.823]

SRDCF [0.822]

MUSTer [0.815]

LCT [0.793]

KCF [0.765]

TGPR [0.751]

DSST [0.742]


Pre

cis

ion

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9


HCF [0.837]

Ours [0.814]

FCNT [0.778]

MEEM [0.769]

MUSTer [0.752]

SRDCF [0.750]

LCT [0.715]

KCF [0.702]

DSST [0.698]

TGPR [0.665]


Pre

cis

ion

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9


HCF [0.854]

Ours [0.852]

MEEM [0.824]

FCNT [0.820]

SRDCF [0.808]

MUSTer [0.782]

LCT [0.774]

KCF [0.749]

TGPR [0.721]

DSST [0.716]


Pre

cis

ion

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9


Ours [0.858]

HCF [0.840]

FCNT [0.825]

SRDCF [0.800]

MEEM [0.787]

MUSTer [0.771]

KCF [0.727]

LCT [0.725]

DSST [0.717]

TGPR [0.676]


Pre

cis

ion

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1occlusion (27)

HCF [0.840]

SRDCF [0.822]

Ours [0.796]

MUSTer [0.773]

MEEM [0.771]

FCNT [0.762]

LCT [0.762]

KCF [0.746]

DSST [0.710]

TGPR [0.684]


Pre

cis

ion

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1deformation (17)

HCF [0.890]

Ours [0.867]

FCNT [0.860]

MEEM [0.859]

SRDCF [0.823]

MUSTer [0.810]

LCT [0.807]

KCF [0.765]

TGPR [0.760]

DSST [0.727]


Pre

cis

ion

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9motion blur (12)

HCF [0.785]

SRDCF [0.756]

Ours [0.745]

FCNT [0.736]

MEEM [0.724]

MUSTer [0.701]

LCT [0.652]

KCF [0.626]

Struck [0.617]

TGPR [0.607]


Pre

cis

ion

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9fast motion (17)

Ours [0.757]

HCF [0.738]

FCNT [0.735]

SRDCF [0.714]

MEEM [0.710]

MUSTer [0.635]

LCT [0.615]

TGPR [0.582]

Struck [0.580]

KCF [0.578]


Pre

cis

ion

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9


HCF [0.847]

Ours [0.847]

FCNT [0.824]

MEEM [0.797]

SRDCF [0.773]

MUSTer [0.755]

LCT [0.751]

KCF [0.724]

DSST [0.720]

TGPR [0.712]


Pre

cis

ion

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9P out of view (6)

SRDCF [0.726]

HCF [0.692]

MEEM [0.692]

FCNT [0.685]

Ours [0.683]

MUSTer [0.671]

KCF [0.643]

LCT [0.621]

DSST [0.587]

TGPR [0.514]


Pre

cis

ion

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9


HCF [0.874]

Ours [0.839]

SRDCF [0.805]

FCNT [0.799]

MEEM [0.793]

KCF [0.776]

MUSTer [0.775]

LCT [0.747]

TGPR [0.721]

DSST [0.697]


Pre

cis

ion

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1low resolution (4)

MEEM [0.946]

Ours [0.941]

FCNT [0.932]

HCF [0.930]

MUSTer [0.798]

SRDCF [0.765]

DSST [0.729]

LCT [0.727]

SCM [0.683]

KCF [0.675]


Success r

ate

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1Success plots of TRE

SRDCF [0.631]

HCF [0.616]

MUSTer [0.614]

FCNT [0.607]

LCT [0.604]

Ours [0.600]

MEEM [0.583]

KCF [0.554]

TGPR [0.547]

DSST [0.536]


Success r

ate

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9


HCF [0.591]

SRDCF [0.575]

FCNT [0.571]

Ours [0.570]

MUSTer [0.567]

MEEM [0.554]

LCT [0.546]

DSST [0.514]

KCF [0.514]

TGPR [0.503]


Success r

ate

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9


SRDCF [0.611]

HCF [0.589]

Ours [0.588]

FCNT [0.582]

LCT [0.579]

MUSTer [0.573]

MEEM [0.570]

KCF [0.528]

TGPR [0.515]

DSST [0.502]


Success r

ate

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9


SRDCF [0.616]

Ours [0.573]

MUSTer [0.565]

FCNT [0.560]

HCF [0.544]

LCT [0.541]

MEEM [0.517]

SCM [0.496]

KCF [0.488]

DSST [0.473]


Success r

ate

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1occlusion (27)

SRDCF [0.632]

HCF [0.600]

LCT [0.584]

MUSTer [0.580]

FCNT [0.561]

MEEM [0.559]

Ours [0.551]

KCF [0.540]

DSST [0.509]

TGPR [0.505]


Success r

ate

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1deformation (17)

HCF [0.647]

SRDCF [0.638]

FCNT [0.627]

LCT [0.625]

MUSTer [0.616]

MEEM [0.613]

Ours [0.607]

KCF [0.573]

TGPR [0.571]

DSST [0.547]


Success r

ate

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1motion blur (12)

HCF [0.594]

SRDCF [0.592]

FCNT [0.560]

MUSTer [0.556]

MEEM [0.553]

Ours [0.550]

LCT [0.523]

KCF [0.493]

Struck [0.485]

TGPR [0.483]


Success r

ate

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1fast motion (17)

SRDCF [0.572]

HCF [0.560]

Ours [0.557]

FCNT [0.553]

MEEM [0.542]

MUSTer [0.518]

LCT [0.503]

Struck [0.464]

TGPR [0.461]

KCF [0.456]


Success r

ate

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9


SRDCF [0.591]

HCF [0.589]

Ours [0.585]

FCNT [0.577]

LCT [0.569]

MUSTer [0.561]

MEEM [0.556]

KCF [0.518]

DSST [0.513]

TGPR [0.511]


Success r

ate

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9out of view (6)

SRDCF [0.591]

MEEM [0.581]

HCF [0.557]

MUSTer [0.555]

FCNT [0.550]

KCF [0.539]

Ours [0.530]

LCT [0.518]

DSST [0.505]

TGPR [0.440]


Success r

ate

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9


HCF [0.631]

SRDCF [0.604]

FCNT [0.583]

Ours [0.581]

MEEM [0.577]

MUSTer [0.574]

KCF [0.565]

LCT [0.557]

TGPR [0.530]

DSST [0.518]


Success r

ate

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1Slow resolution (4)

Ours [0.530]

SRDCF [0.505]

FCNT [0.493]

SCM [0.463]

HCF [0.456]

MEEM [0.454]

MUSTer [0.450]

LCT [0.441]

DSST [0.380]

DLT [0.364]

Figure 7. Detailed Results of TRE on OTB-51 benchmark.


Pre

cis

ion

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1Precision plots of SRE

HCF [0.849]

Ours [0.848]

MUSTer [0.822]

FCNT [0.817]

SRDCF [0.799]

MEEM [0.771]

LCT [0.766]

DSST [0.702]

TGPR [0.693]

KCF [0.683]


Pre

cis

ion

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1 illumination variation (23)

Ours [0.834]

HCF [0.794]

FCNT [0.741]

MUSTer [0.736]

SRDCF [0.691]

MEEM [0.685]

LCT [0.675]

DSST [0.654]

KCF [0.623]

TGPR [0.600]


Pre

cis

ion

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9


Ours [0.838]

HCF [0.828]

MUSTer [0.797]

FCNT [0.784]

LCT [0.776]

MEEM [0.775]

SRDCF [0.763]

DSST [0.702]

KCF [0.691]

TGPR [0.666]


Pre

cis

ion

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9


Ours [0.845]

HCF [0.832]

FCNT [0.799]

MUSTer [0.769]

SRDCF [0.747]

MEEM [0.732]

DSST [0.696]

LCT [0.693]

KCF [0.632]

Struck [0.607]


Pre

cis

ion

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1occlusion (27)

HCF [0.814]

MUSTer [0.798]

SRDCF [0.774]

Ours [0.767]

LCT [0.755]

FCNT [0.734]

MEEM [0.729]

DSST [0.700]

KCF [0.697]

TGPR [0.643]


Pre

cis

ion

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1deformation (17)

Ours [0.848]

FCNT [0.843]

HCF [0.836]

LCT [0.810]

MUSTer [0.800]

SRDCF [0.773]

MEEM [0.759]

KCF [0.733]

TGPR [0.712]

DSST [0.671]


Pre

cis

ion

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1motion blur (12)

HCF [0.807]

Ours [0.786]

SRDCF [0.723]

FCNT [0.717]

MEEM [0.691]

MUSTer [0.688]

LCT [0.608]

Struck [0.587]

KCF [0.567]

TGPR [0.561]


Pre

cis

ion

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1fast motion (17)

Ours [0.788]

HCF [0.748]

FCNT [0.700]

SRDCF [0.697]

MEEM [0.694]

MUSTer [0.658]

LCT [0.591]

Struck [0.577]

KCF [0.545]

TGPR [0.544]


Pre

cis

ion

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9


HCF [0.839]

Ours [0.832]

FCNT [0.791]

MUSTer [0.768]

MEEM [0.752]

LCT [0.747]

SRDCF [0.730]

DSST [0.704]

KCF [0.667]

TGPR [0.648]


Pre

cis

ion

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1out of view (6)

Ours [0.703]

MEEM [0.690]

SRDCF [0.661]

MUSTer [0.655]

HCF [0.644]

FCNT [0.616]

LCT [0.612]

KCF [0.533]

DSST [0.504]

TLD [0.463]


Pre

cis

ion

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9


Ours [0.865]

HCF [0.851]

MUSTer [0.776]

SRDCF [0.737]

FCNT [0.735]

MEEM [0.734]

LCT [0.732]

TGPR [0.698]

KCF [0.693]

DSST [0.655]


Pre

cis

ion

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1low resolution (4)

FCNT [0.952]

MEEM [0.915]

HCF [0.876]

Ours [0.843]

SRDCF [0.759]

MUSTer [0.679]

LCT [0.668]

DSST [0.633]

SCM [0.627]

DLT [0.576]


Success r

ate

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1Success plots of SRE

SRDCF [0.570]

MUSTer [0.567]

FCNT [0.555]

Ours [0.559]

HCF [0.551]

LCT [0.534]

MEEM [0.521]

TGPR [0.475]

KCF [0.462]

DSST [0.457]


Success r

ate

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9


Ours [0.552]

FCNT [0.528]

MUSTer [0.521]

HCF [0.514]

SRDCF [0.500]

MEEM [0.486]

LCT [0.476]

DSST [0.438]

KCF [0.435]

TGPR [0.428]


Success r

ate

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9


Ours [0.549]

SRDCF [0.543]

MUSTer [0.541]

LCT [0.540]

FCNT [0.537]

HCF [0.534]

MEEM [0.514]

KCF [0.461]

DSST [0.452]

TGPR [0.452]


Success r

ate

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9


SRDCF [0.539]

Ours [0.538]

MUSTer [0.522]

FCNT [0.522]

HCF [0.492]

LCT [0.477]

MEEM [0.463]

SCM [0.438]

DSST [0.413]

DLT [0.402]


Success r

ate

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1occlusion (27)

SRDCF [0.557]

MUSTer [0.553]

HCF [0.540]

LCT [0.535]

FCNT [0.522]

Ours [0.515]

MEEM [0.510]

KCF [0.468]

DSST [0.452]

TGPR [0.445]


Success r

ate

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1deformation (17)

FCNT [0.588]

LCT [0.584]

MUSTer [0.580]

Ours [0.575]

SRDCF [0.566]

HCF [0.564]

MEEM [0.516]

KCF [0.508]

TGPR [0.503]

DSST [0.463]


Success r

ate

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1motion blur (12)

HCF [0.565]

Ours [0.547]

SRDCF [0.535]

FCNT [0.524]

MEEM [0.513]

MUSTer [0.510]

LCT [0.459]

Struck [0.452]

KCF [0.425]

TGPR [0.420]


Success r

ate

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1fast motion (17)

Ours [0.553]

HCF [0.534]

MEEM [0.518]

FCNT [0.517]

SRDCF [0.516]

MUSTer [0.490]

Struck [0.451]

LCT [0.451]

KCF [0.415]

TGPR [0.412]


Success r

ate

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9


HCF [0.537]

Ours [0.536]

FCNT [0.530]

MUSTer [0.525]

SRDCF [0.520]

LCT [0.518]

MEEM [0.494]

DSST [0.460]

KCF [0.450]

TGPR [0.438]


Success r

ate

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1out of view (6)

MEEM [0.575]

Ours [0.528]

HCF [0.526]

MUSTer [0.517]

SRDCF [0.516]

FCNT [0.502]

LCT [0.489]

KCF [0.455]

DSST [0.426]

Struck [0.421]


Success r

ate

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9


HCF [0.569]

Ours [0.555]

MUSTer [0.542]

SRDCF [0.527]

MEEM [0.517]

FCNT [0.513]

LCT [0.512]

TGPR [0.483]

KCF [0.483]

DSST [0.451]


Success r

ate

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1low resolution (4)

SRDCF [0.472]

FCNT [0.469]

Ours [0.449]

MEEM [0.396]

HCF [0.395]

SCM [0.385]

MUSTer [0.358]

LCT [0.349]

DLT [0.337]

DSST [0.299]

Figure 8. Detailed Results of SRE on OTB-51 benchmark.


Pre

cis

ion

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1Precision plots of OPE

Ours [0.851]

HCF [0.837]

FCNT [0.798]

SRDCF [0.789]

MEEM [0.781]

MUSTer [0.774]

LCT [0.762]

DSST [0.695]

KCF [0.692]

TGPR [0.643]


Pre

cis

ion

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9


Ours [0.875]

HCF [0.806]

SRDCF [0.774]

MUSTer [0.767]

FCNT [0.766]

MEEM [0.735]

LCT [0.729]

DSST [0.709]

KCF [0.697]

TGPR [0.617]


Pre

cis

ion

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9


Ours [0.865]

FCNT [0.811]

HCF [0.809]

MEEM [0.808]

LCT [0.762]

MUSTer [0.760]

SRDCF [0.740]

DSST [0.705]

KCF [0.694]

TGPR [0.655]


Pre

cis

ion

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9


Ours [0.834]

HCF [0.805]

FCNT [0.773]

MEEM [0.757]

SRDCF [0.748]

MUSTer [0.729]

LCT [0.699]

DSST [0.681]

KCF [0.641]

Struck [0.614]


Pre

cis

ion

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9occlusion (44)

Ours [0.805]

HCF [0.763]

MEEM [0.760]

MUSTer [0.751]

FCNT [0.738]

SRDCF [0.726]

LCT [0.694]

DSST [0.657]

KCF [0.651]

TGPR [0.610]


Pre

cis

ion

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9deformation (39)

Ours [0.801]

HCF [0.790]

MEEM [0.777]

FCNT [0.771]

SRDCF [0.724]

MUSTer [0.703]

LCT [0.703]

TGPR [0.653]

KCF [0.649]

DSST [0.609]


Pre

cis

ion

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9 motion blur (29)

HCF [0.804]

Ours [0.781]

SRDCF [0.767]

FCNT [0.735]

MEEM [0.731]

MUSTer [0.678]

LCT [0.669]

DSST [0.611]

KCF [0.600]

Struck [0.585]


Pre

cis

ion

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1 fast motion (37)

Ours [0.829]

HCF [0.827]

SRDCF [0.781]

MEEM [0.781]

FCNT [0.752]

MUSTer [0.717]

LCT [0.715]

Struck [0.639]

KCF [0.630]

DSST [0.612]


Pre

cis

ion

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9


Ours [0.876]

HCF [0.854]

FCNT [0.830]

MEEM [0.794]

LCT [0.782]

MUSTer [0.773]

SRDCF [0.745]

DSST [0.724]

KCF [0.693]

TGPR [0.659]


Pre

cis

ion

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9 out of view (14)

Ours [0.734]

MEEM [0.685]

HCF [0.677]

FCNT [0.629]

SRDCF [0.597]

LCT [0.592]

MUSTer [0.591]

DLT [0.558]

KCF [0.498]

TGPR [0.493]


Pre

cis

ion

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9


Ours [0.903]

HCF [0.843]

MUSTer [0.784]

SRDCF [0.775]

FCNT [0.750]

MEEM [0.746]

LCT [0.734]

KCF [0.712]

DSST [0.702]

SCM [0.603]


Pre

cis

ion

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1low resolution (9)

Ours [0.918]

HCF [0.847]

FCNT [0.845]

MEEM [0.808]

SRDCF [0.765]

SCM [0.761]

DLT [0.751]

MUSTer [0.747]

DSST [0.708]

LCT [0.699]


Success r

ate

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1Success plots of OPE

SRDCF [0.598]

MUSTer [0.577]

Ours [0.572]

HCF [0.562]

LCT [0.562]

FCNT [0.555]

MEEM [0.530]

DSST [0.475]

KCF [0.475]

Struck [0.458]


Success r

ate

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9


SRDCF [0.599]

MUSTer [0.592]

Ours [0.587]

LCT [0.556]

HCF [0.540]

FCNT [0.536]

MEEM [0.521]

DSST [0.486]

SCM [0.478]

KCF [0.474]


Success r

ate

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9


Ours [0.571]

FCNT [0.551]

LCT [0.547]

MUSTer [0.546]

SRDCF [0.544]

HCF [0.531]

MEEM [0.530]

DSST [0.465]

KCF [0.463]

TGPR [0.456]


Success r

ate

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1 scale variation (61)

SRDCF [0.561]

Ours [0.532]

MUSTer [0.524]

FCNT [0.506]

LCT [0.500]

HCF [0.486]

MEEM [0.479]

SCM [0.447]

DSST [0.414]

Struck [0.407]


Success r

ate

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1occlusion (44)

MUSTer [0.567]

SRDCF [0.550]

Ours [0.544]

HCF [0.520]

LCT [0.515]

FCNT [0.514]

MEEM [0.512]

KCF [0.456]

DSST [0.446]

SCM [0.441]


Success r

ate

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1 deformation (39)

Ours [0.538]

MUSTer [0.534]

SRDCF [0.531]

FCNT [0.529]

HCF [0.525]

LCT [0.507]

MEEM [0.496]

TGPR [0.460]

KCF [0.455]

DSST [0.433]


Success r

ate

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1motion blur (29)

SRDCF [0.594]

HCF [0.585]

Ours [0.571]

FCNT [0.562]

MEEM [0.556]

MUSTer [0.544]

LCT [0.533]

DSST [0.467]

KCF [0.459]

Struck [0.459]


Success r

ate

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1fast motion (37)

SRDCF [0.601]

HCF [0.578]

Ours [0.577]

LCT [0.560]

MUSTer [0.557]

MEEM [0.557]

FCNT [0.551]

Struck [0.472]

KCF [0.465]

DSST [0.452]


Success r

ate

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9


Ours [0.576]

HCF [0.559]

FCNT [0.558]

LCT [0.557]

MUSTer [0.551]

SRDCF [0.544]

MEEM [0.529]

DSST [0.485]

KCF [0.465]

TGPR [0.462]


Success r

ate

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8


Ours [0.515]

MEEM [0.488]

HCF [0.474]

FCNT [0.470]

MUSTer [0.469]

SRDCF [0.460]

LCT [0.452]

KCF [0.393]

DLT [0.384]

DSST [0.374]


Success r

ate

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1 background clutter (31)

Ours [0.593]

HCF [0.585]

SRDCF [0.583]

MUSTer [0.581]

LCT [0.550]

FCNT [0.529]

MEEM [0.519]

KCF [0.497]

DSST [0.477]

SCM [0.476]


Success r

ate

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1low resolution (9)

SRDCF [0.514]

SCM [0.478]

DLT [0.465]

Ours [0.463]

FCNT [0.421]

MUSTer [0.415]

LCT [0.399]

HCF [0.388]

MEEM [0.382]

TLD [0.346]

Figure 9. Detailed Results of OPE on OTB-100 benchmark.


Pre

cis

ion

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1Precision plots of TRE

HCF [0.835]

Ours [0.828]

FCNT [0.813]

MEEM [0.789]

SRDCF [0.781]

MUSTer [0.764]

LCT [0.740]

DSST [0.717]

KCF [0.716]

TGPR [0.697]


Pre

cis

ion

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9


HCF [0.832]

Ours [0.818]

FCNT [0.799]

MEEM [0.783]

SRDCF [0.768]

MUSTer [0.742]

LCT [0.720]

KCF [0.701]

DSST [0.699]

TGPR [0.665]


Pre

cis

ion

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9


Ours [0.830]

HCF [0.822]

FCNT [0.813]

MEEM [0.787]

SRDCF [0.743]

MUSTer [0.732]

LCT [0.716]

KCF [0.693]

TGPR [0.689]

DSST [0.689]


Pre

cis

ion

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9


Ours [0.824]

FCNT [0.807]

HCF [0.804]

MEEM [0.754]

SRDCF [0.751]

MUSTer [0.719]

DSST [0.687]

LCT [0.678]

KCF [0.668]

TGPR [0.647]


Pre

cis

ion

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1occlusion (44)

HCF [0.792]

Ours [0.777]

FCNT [0.756]

MEEM [0.752]

SRDCF [0.751]

MUSTer [0.734]

LCT [0.684]

DSST [0.662]

KCF [0.660]

TGPR [0.644]


Pre

cis

ion

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1deformation (39)

HCF [0.836]

Ours [0.823]

FCNT [0.820]

MEEM [0.791]

SRDCF [0.762]

MUSTer [0.734]

LCT [0.717]

TGPR [0.707]

DSST [0.679]

KCF [0.679]


Pre

cis

ion

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9motion blur (29)

HCF [0.779]

MEEM [0.752]

SRDCF [0.750]

Ours [0.746]

FCNT [0.732]

MUSTer [0.700]

LCT [0.658]

Struck [0.620]

DSST [0.615]

KCF [0.611]


Pre

cis

ion

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1fast motion (37)

HCF [0.786]

Ours [0.785]

FCNT [0.760]

MEEM [0.748]

SRDCF [0.748]

MUSTer [0.690]

LCT [0.642]

Struck [0.627]

KCF [0.624]

DSST [0.600]


Pre

cis

ion

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9


Ours [0.842]

FCNT [0.831]

HCF [0.827]

MEEM [0.769]

MUSTer [0.730]

SRDCF [0.726]

LCT [0.712]

DSST [0.690]

KCF [0.685]

TGPR [0.684]


Pre

cis

ion

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9out of view (14)

Ours [0.655]

FCNT [0.636]

MEEM [0.630]

HCF [0.624]

MUSTer [0.606]

SRDCF [0.597]

DSST [0.520]

KCF [0.509]

LCT [0.504]

TGPR [0.499]


Pre

cis

ion

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9


HCF [0.856]

Ours [0.838]

SRDCF [0.814]

MEEM [0.798]

FCNT [0.791]

KCF [0.783]

MUSTer [0.779]

LCT [0.747]

DSST [0.735]

TGPR [0.693]


Pre

cis

ion

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1low resolution (9)

Ours [0.926]

FCNT [0.905]

HCF [0.885]

MEEM [0.879]

SRDCF [0.764]

TGPR [0.757]

SCM [0.752]

DSST [0.748]

MUSTer [0.745]

Struck [0.733]


Success r

ate

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1Success plots of TRE

SRDCF [0.608]

HCF [0.592]

FCNT [0.588]

Ours [0.579]

MUSTer [0.579]

LCT [0.567]

MEEM [0.565]

KCF [0.522]

DSST [0.520]

TGPR [0.513]


Success r

ate

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9


SRDCF [0.605]

HCF [0.600]

FCNT [0.589]

Ours [0.581]

MEEM [0.577]

MUSTer [0.576]

LCT [0.557]

DSST [0.525]

KCF [0.516]

TGPR [0.502]


Success r

ate

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9


FCNT [0.574]

Ours [0.572]

SRDCF [0.566]

HCF [0.563]

MEEM [0.546]

LCT [0.536]

MUSTer [0.533]

TGPR [0.490]

KCF [0.489]

DSST [0.479]


Success r

ate

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9


SRDCF [0.583]

Ours [0.560]

FCNT [0.557]

HCF [0.533]

MUSTer [0.528]

MEEM [0.513]

LCT [0.510]

DSST [0.463]

KCF [0.457]

TGPR [0.455]


Success r

ate

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1occlusion (44)

SRDCF [0.583]

HCF [0.558]

FCNT [0.546]

MUSTer [0.545]

Ours [0.540]

MEEM [0.534]

LCT [0.524]

KCF [0.481]

DSST [0.470]

TGPR [0.466]


Success r

ate

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1deformation (39)

HCF [0.588]

FCNT [0.586]

SRDCF [0.582]

Ours [0.576]

MEEM [0.558]

MUSTer [0.551]

LCT [0.542]

TGPR [0.516]

KCF [0.504]

DSST [0.501]


Success r

ate

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1motion blur (29)

SRDCF [0.595]

HCF [0.591]

MEEM [0.579]

FCNT [0.573]

Ours [0.564]

MUSTer [0.555]

LCT [0.535]

Struck [0.501]

DSST [0.482]

KCF [0.476]


Success r

ate

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1fast motion (37)

SRDCF [0.598]

HCF [0.581]

Ours [0.575]

FCNT [0.571]

MEEM [0.566]

MUSTer [0.549]

LCT [0.522]

Struck [0.494]

KCF [0.482]

TGPR [0.478]


Success r

ate

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9


Ours [0.580]

FCNT [0.579]

HCF [0.572]

SRDCF [0.559]

MEEM [0.544]

LCT [0.539]

MUSTer [0.537]

TGPR [0.496]

KCF [0.493]

DSST [0.491]


Success r

ate

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9out of view (14)

FCNT [0.491]

MEEM [0.478]

Ours [0.478]

SRDCF [0.471]

MUSTer [0.465]

HCF [0.459]

LCT [0.408]

KCF [0.399]

DSST [0.397]

TGPR [0.388]


Success r

ate

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9


SRDCF [0.623]

HCF [0.619]

MUSTer [0.587]

Ours [0.584]

MEEM [0.581]

FCNT [0.578]

KCF [0.570]

LCT [0.567]

DSST [0.543]

TGPR [0.512]


Success r

ate

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1low resolution (9)

Ours [0.523]

SRDCF [0.511]

FCNT [0.499]

SCM [0.491]

HCF [0.452]

MEEM [0.446]

LCT [0.439]

DLT [0.430]

TGPR [0.424]

MUSTer [0.418]

Figure 10. Detailed Results of TRE on OTB-100 benchmark.


Pre

cis

ion

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1Precision plots of SRE

HCF [0.800]

Ours [0.793]

FCNT [0.759]

SRDCF [0.756]

MUSTer [0.741]

MEEM [0.731]

LCT [0.701]

DSST [0.662]

KCF [0.640]

TGPR [0.626]


Pre

cis

ion

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9


Ours [0.823]

HCF [0.792]

FCNT [0.743]

SRDCF [0.726]

MUSTer [0.721]

MEEM [0.711]

LCT [0.670]

DSST [0.666]

KCF [0.642]

TGPR [0.556]


Pre

cis

ion

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9


Ours [0.792]

HCF [0.774]

FCNT [0.771]

MEEM [0.749]

MUSTer [0.726]

SRDCF [0.712]

LCT [0.708]

DSST [0.649]

TGPR [0.641]

KCF [0.631]


Pre

cis

ion

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9scale variation (61)

HCF [0.767]

Ours [0.765]

FCNT [0.741]

SRDCF [0.720]

MEEM [0.699]

MUSTer [0.694]

LCT [0.644]

DSST [0.637]

Struck [0.595]

KCF [0.589]


Pre

cis

ion

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9occlusion (44)

HCF [0.731]

Ours [0.718]

MUSTer [0.712]

MEEM [0.710]

FCNT [0.702]

SRDCF [0.692]

LCT [0.643]

DSST [0.611]

TGPR [0.589]

KCF [0.581]


Pre

cis

ion

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9deformation (39)

HCF [0.754]

FCNT [0.745]

Ours [0.742]

MEEM [0.688]

SRDCF [0.687]

LCT [0.667]

MUSTer [0.659]

TGPR [0.623]

KCF [0.604]

DSST [0.577]


Pre

cis

ion

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9motion blur (29)

HCF [0.752]

Ours [0.712]

SRDCF [0.707]

MEEM [0.694]

FCNT [0.670]

MUSTer [0.644]

Struck [0.602]

LCT [0.581]

DSST [0.540]

TGPR [0.534]


Pre

cis

ion

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1 fast motion (37)

HCF [0.781]

Ours [0.772]

SRDCF [0.737]

FCNT [0.704]

MEEM [0.703]

MUSTer [0.676]

LCT [0.628]

Struck [0.624]

TGPR [0.577]

KCF [0.576]


Pre

cis

ion

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9


HCF [0.821]

Ours [0.800]

FCNT [0.784]

MEEM [0.748]

MUSTer [0.736]

LCT [0.724]

SRDCF [0.713]

DSST [0.668]

TGPR [0.645]

KCF [0.632]


Pre

cis

ion

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8out of view (14)

Ours [0.646]

MEEM [0.619]

HCF [0.592]

FCNT [0.583]

SRDCF [0.580]

MUSTer [0.570]

LCT [0.498]

Struck [0.468]

DSST [0.463]

TGPR [0.451]


Pre

cis

ion

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9background clutter (31)

Ours [0.815]

HCF [0.803]

MUSTer [0.738]

SRDCF [0.729]

FCNT [0.708]

MEEM [0.682]

DSST [0.670]

LCT [0.670]

KCF [0.662]

TGPR [0.587]


Pre

cis

ion

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1 low resolution (9)

FCNT [0.877]

Ours [0.827]

HCF [0.819]

MEEM [0.811]

SRDCF [0.775]

DLT [0.726]

MUSTer [0.696]

SCM [0.692]

LCT [0.691]

Struck [0.685]


Success r

ate

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1Success plots of SRE

SRDCF [0.543]

FCNT [0.519]

Ours [0.519]

MUSTer [0.516]

HCF [0.515]

MEEM [0.492]

LCT [0.488]

TGPR [0.433]

DSST [0.433]

KCF [0.432]


Success r

ate

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9


Ours [0.539]

SRDCF [0.532]

FCNT [0.519]

MUSTer [0.513]

HCF [0.502]

MEEM [0.489]

LCT [0.471]

DSST [0.432]

KCF [0.426]

SCM [0.401]


Success r

ate

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9


FCNT [0.515]

Ours [0.513]

SRDCF [0.506]

HCF [0.496]

MEEM [0.494]

MUSTer [0.493]

LCT [0.489]

TGPR [0.429]

DSST [0.419]

KCF [0.418]


Success r

ate

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9


SRDCF [0.516]

Ours [0.488]

FCNT [0.482]

MUSTer [0.471]

HCF [0.456]

MEEM [0.444]

LCT [0.442]

Struck [0.393]

SCM [0.385]

DSST [0.384]


Success r

ate

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9occlusion (44)

SRDCF [0.502]

MUSTer [0.496]

FCNT [0.491]

HCF [0.480]

MEEM [0.480]

Ours [0.477]

LCT [0.456]

TGPR [0.400]

DSST [0.400]

KCF [0.398]


Success r

ate

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9deformation (39)

FCNT [0.504]

Ours [0.494]

SRDCF [0.488]

HCF [0.483]

MUSTer [0.473]

LCT [0.463]

MEEM [0.453]

TGPR [0.431]

KCF [0.414]

DSST [0.394]


Success r

ate

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9motion blur (29)

HCF [0.535]

SRDCF [0.529]

MEEM [0.523]

Ours [0.519]

FCNT [0.509]

MUSTer [0.487]

Struck [0.471]

LCT [0.453]

TGPR [0.415]

TLD [0.409]


Success r

ate

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1fast motion (37)

SRDCF [0.545]

Ours [0.544]

HCF [0.534]

MEEM [0.508]

FCNT [0.507]

MUSTer [0.498]

LCT [0.473]

Struck [0.467]

TGPR [0.430]

KCF [0.425]


Success r

ate

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9


HCF [0.524]

FCNT [0.522]

Ours [0.518]

SRDCF [0.505]

MUSTer [0.501]

LCT [0.498]

MEEM [0.498]

TGPR [0.439]

DSST [0.438]

KCF [0.427]


Success r

ate

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8


Ours [0.460]

MEEM [0.457]

FCNT [0.436]

HCF [0.432]

MUSTer [0.431]

SRDCF [0.430]

LCT [0.378]

Struck [0.371]

DSST [0.346]

TLD [0.343]


Success r

ate

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9


HCF [0.527]

Ours [0.526]

SRDCF [0.524]

MUSTer [0.513]

FCNT [0.488]

MEEM [0.469]

LCT [0.464]

KCF [0.446]

DSST [0.442]

SCM [0.405]


Success r

ate

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9low resolution (9)

SRDCF [0.492]

FCNT [0.441]

Ours [0.420]

DLT [0.419]

SCM [0.412]

LCT [0.374]

MEEM [0.373]

MUSTer [0.371]

HCF [0.362]

TGPR [0.331]

Figure 11. Detailed Results of SRE on OTB-100 benchmark.

Table 2. Overall AR ranking on VOt2016.

Method Accuracy RobustnessACT 19.75 26.79ANT 14.11 25.21

ART DSST 13.89 27.25ASMS 12.18 26.61BDF 29.35 34.98BST 33.11 20.09

CCCT 21.05 19.18CCOT 8.11 10.00CDTT 21.75 22.72CMT 31.93 56.77CTF 13.96 48.09

ColorKCF 11.72 22.68DAT 15.84 21.35DDC 5.84 14.86DFST 15.54 30.44DFT 18.47 41.53DNT 10.09 15.56DPT 12.61 21.60

DPTG 13.95 25.33DSST2014 11.26 27.30

DeepSRDCF 10.47 16.02EBT 17.39 7.96FCF 3.96 16.58FCT 29.09 34.23FRT 31.07 48.54FoT 29.21 35.96GCF 8.25 19.72

GGTv2 7.77 21.89HCF 12.11 14.02

HMMTxD 8.39 23.44HT 26.67 35.96IVT 24.74 42.95

KCF2014 13.30 22.39KCF SMXPC 6.84 17.53

LGT 25.33 29.89LT FLO 21.40 51.51

Method Accuracy RobustnessLoFT Lite 40.68 51.33

MAD 12.84 18.65MDNet N 5.58 13.02

MIL 23.47 33.93MLDF 13.44 8.07MWCF 13.49 21.26 6

MatFlow 26.37 32.14Matrioska 22.68 45.09NSAMF 10.77 15.25OEST 8.98 25.72Ours 26.23 40.60

PKLTF 25.23 31.70RFD CF2 12.58 12.42

SAMF2014 9.51 19.82SCT4 15.35 21.60SHCT 5.39 15.74SMPR 18.51 32.51SODLT 10.46 21.42SRBT 11.79 14.70

SRDCF 7.39 17.49SSAT 3.61 12.40

SSKCF 7.98 15.65STAPLEp 4.28 14.84

STC 31.61 40.72STRUCK2014 21.35 38.72

SWCF 13.96 24.81SiamAN 10.35 18.81SiamRN 3.81 19.11Staple 6.07 16.33TCNN 5.49 12.72TGPR 16.65 28.72

TricTRACK 19.09 25.70deepMKCF 6.91 16.49

ncc 15.82 56.82sKCF 14.28 34.81

supplementary ﬁle for robust visual tracking using oblique ......for the cases with a large number...

Documents