Download textbook Advances In Multimedia Information Processing Pcm 2016 17Th Pacific Rim Conference On Multimedia Xi An China September 15 16 2016 Proceedings Part I 1St Edition Enqing Chen ebook all chapter pdf
Download textbook Advances In Multimedia Information Processing Pcm 2016 17Th Pacific Rim Conference On Multimedia Xi An China September 15 16 2016 Proceedings Part I 1St Edition Enqing Chen ebook all chapter pdf
Download textbook Advances In Multimedia Information Processing Pcm 2016 17Th Pacific Rim Conference On Multimedia Xi An China September 15 16 2016 Proceedings Part I 1St Edition Enqing Chen ebook all chapter pdf
https://textbookfull.com/product/multimedia-modeling-22nd-
international-conference-mmm-2016-miami-fl-usa-
january-4-6-2016-proceedings-part-i-1st-edition-qi-tian/
https://textbookfull.com/product/web-age-information-
management-17th-international-conference-waim-2016-nanchang-
china-june-3-5-2016-proceedings-part-i-1st-edition-bin-cui/
https://textbookfull.com/product/neural-information-
processing-23rd-international-conference-iconip-2016-kyoto-japan-
october-16-21-2016-proceedings-part-iv-1st-edition-akira-hirose/
Advances in Multimedia
LNCS 9916
Information Processing –
PCM 2016
17th Pacific-Rim Conference on Multimedia
Xi‘an, China, September 15–16, 2016
Proceedings, Part I
123
Lecture Notes in Computer Science 9916
Commenced Publication in 1973
Founding and Former Series Editors:
Gerhard Goos, Juris Hartmanis, and Jan van Leeuwen
Editorial Board
David Hutchison
Lancaster University, Lancaster, UK
Takeo Kanade
Carnegie Mellon University, Pittsburgh, PA, USA
Josef Kittler
University of Surrey, Guildford, UK
Jon M. Kleinberg
Cornell University, Ithaca, NY, USA
Friedemann Mattern
ETH Zurich, Zurich, Switzerland
John C. Mitchell
Stanford University, Stanford, CA, USA
Moni Naor
Weizmann Institute of Science, Rehovot, Israel
C. Pandu Rangan
Indian Institute of Technology, Madras, India
Bernhard Steffen
TU Dortmund University, Dortmund, Germany
Demetri Terzopoulos
University of California, Los Angeles, CA, USA
Doug Tygar
University of California, Berkeley, CA, USA
Gerhard Weikum
Max Planck Institute for Informatics, Saarbrücken, Germany
More information about this series at http://www.springer.com/series/7409
Enqing Chen Yihong Gong
•
Advances in Multimedia
Information Processing –
PCM 2016
17th Pacific-Rim Conference on Multimedia
Xi’an, China, September 15–16, 2016
Proceedings, Part I
123
Editors
Enqing Chen Yun Tie
Zhengzhou University Zhengzhou University
Zhengzhou Zhengzhou
China China
Yihong Gong
Jiaotong University
Xi’an
China
LNCS Sublibrary: SL3 – Information Systems and Applications, incl. Internet/Web, and HCI
The 17th Pacific-Rim Conference on Multimedia (PCM 2016) was held in Xi’an,
China, during September 15–16, 2016, and hosted by the Xi’an Jiaotong University
(XJTU). PCM is a leading international conference for researchers and industry
practitioners to share their new ideas, original research results, and practical devel-
opment experiences from all multimedia-related areas.
It was a great honor for XJTU to host PCM 2016, one of the most longstanding
multimedia conferences, in Xi’an, China. Xi’an Jiaotong University, located in the
capital of Shaanxi province, is one of the key universities run by the Ministry of
Education, China. Recently its multimedia-related research has been attracting
increasing attention from the local and international multimedia community. For over
2000 years, Xi’an has been the center for political and economic developments and the
capital city of many Chinese dynasties, with the richest cultural and historical heritage,
including the world-famous Terracotta Warriors, Big Wild Goose Pagoda, etc. We
hope that our venue made PCM 2016 a memorable experience for all participants.
PCM 2016 featured a comprehensive program. The 202 submissions from authors
of more than ten countries included a large number of high-quality papers in multi-
media content analysis, multimedia signal processing and communications, and mul-
timedia applications and services. We thank our 28 Technical Program Committee
members who spent many hours reviewing papers and providing valuable feedback to
the authors. From the total of 202 submissions to the main conference and based on at
least three reviews per submission, the program chairs decided to accept 111 regular
papers (54 %) among which 67 were posters (33 %). This volume of the conference
proceedings contains the abstracts of two invited talks and all the regular, poster, and
special session papers.
The technical program is an important aspect but only achieves its full impact if
complemented by challenging keynotes. We are extremely pleased and grateful to have
had two exceptional keynote speakers, Wen Gao and Alex Hauptmann, accept our
invitation and present interesting ideas and insights at PCM 2016.
We are also heavily indebted to many individuals for their significant contributions.
We thank the PCM Steering Committee for their invaluable input and guidance on
crucial decisions. We wish to acknowledge and express our deepest appreciation to the
honorary chairs, Nanning Zheng, Shin’chi Satoh, general chairs, Yihong Gong, Tho-
mas Plagemann, Ke Lu, Jianping Fan, program chairs, Meng Wang, Qi Tian, Abdul-
motaleb EI Saddik, Yun Tie, organizing chairs, Jinye Peng, Xinbo Gao, Ziyu Guan,
Yizhou Wang, publicity chairs, Xueming Qian, Xiaojiang Chen, Cheng Jin, Xiangyang
Xue, publication chairs, Jun Wu, Enqing Chen, local Arrangements Chairs, Kuizi Mei,
Xuguang Lan, special session chairs, Jianbing Shen, Jialie Shen, Jianru Xue, demo
chairs, Yugang Jiang, Jitao Sang, finance and registration chair, Shuchan Gao. Without
their efforts and enthusiasm, PCM 2016 would not have become a reality. Moreover,
we want to thank our sponsors: Springer, Peking University, Zhengzhou University,
VI Preface
Honorary Chairs
Nanning Zheng Xi’an Jiaotong University, China
Shin’chi Satoh National Institute of Informatics, Japan
General Chairs
Yihong Gong Xi’an Jiaotong University, China
Thomas Plagemann University of Oslo, Norway
Ke Lu University of Chinese Academy of Sciences, China
Jianping Fan University of North Carolina at Charlotte, USA
Program Chairs
Meng Wang Hefei University of Technology, China
Qi Tian University of Texas at San Antonio, USA
Abdulmotaleb EI Saddik University of Ottawa, Canada
Yun Tie Zhengzhou University, China
Organizing Chairs
Jinye Peng Northwest University, China
Xinbo Gao Xidian University, China
Ziyu Guan Northwest University, China
Yizhou Wang Peking University, China
Publicity Chairs
Xueming Qian Xi’an Jiaotong University, China
Xiaojiang Chen Northwest University, China
Cheng Jin Fudan University, China
Xiangyang Xue Fudan University, China
Publication Chairs
Jun Wu Northwestern Polytechnical University, China
Enqing Chen Zhengzhou University, China
VIII Organization
Demo Chairs
Yugang Jiang Fudan University, China
Jitao Sang Institute of Automation, Chinese Academy of Sciences,
China
Multi-scale Point Set Saliency Detection Based on Site Entropy Rate . . . . . . 366
Yu Guo, Fei Wang, Pengyu Liu, Jingmin Xin, and Nanning Zheng
Deep Metric Learning with Improved Triplet Loss for Face Clustering
in Videos . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 497
Shun Zhang, Yihong Gong, and Jinjun Wang
Sparse Matrix Based Hashing for Approximate Nearest Neighbor Search . . . . 559
Min Wang, Wengang Zhou, Qi Tian, and Houqiang Li
Dynamic Contour Matching for Lossy Screen Content Picture Intra Coding . . . 326
Hu Yuan, Tao Pin, and Yuanchun Shi
Product Image Search with Deep Attribute Mining and Re-ranking . . . . . . . . 561
Xin Zhou, Yuqi Zhang, Xiuxiu Bai, Jihua Zhu, Li Zhu, and Xueming Qian
A New Rate Control Algorithm Based on Region of Interest for HEVC . . . . 571
Liquan Shen, Qianqian Hu, Zhi Liu, and Ping An
GIP: Generic Image Prior for No Reference Image Quality Assessment . . . . . 600
Qingbo Wu, Hongliang Li, and King N. Ngan
Social Media Profiler: Inferring Your Social Media Personality from Visual
Attributes in Portrait . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 640
Jie Nie, Lei Huang, Peng Cui, Zhen Li, Yan Yan, Zhiqiang Wei,
and Wenwu Zhu
Scene Parsing with Deep Features and Spatial Structure Learning . . . . . . . . . 715
Hui Yu, Yuecheng Song, Wenyu Ju, and Zhenbao Liu
1 Introduction
In computer vision field, object tracking plays a crucial role for its various appli-
cations, such as surveillance and robotics [14]. To develop a robust tracker,
numerous algorithms have been proposed. Despite reasonable good results of
these methods, visual tracking remains a challenge due to appearance variations
caused by occlusion and deformation. To address these problems, a wide range
of appearance models have been presented. In general, these appearance models
can be categorized into two types: discriminative models [2,3,6,9,10,13,18,20]
and generative models [1,5,7,8,15,16].
Discriminative algorithms focus on building online classifiers to distinguish
the target from the background. These methods employ both the foreground
and background information. In [2], an adaptive ensemble of classifier is trained
to separate target pixels from background pixels. Kalal et al. [13] introduce a
P-N learning algorithm for object tracking. However, this tracking method easily
c Springer International Publishing AG 2016
E. Chen et al. (Eds.): PCM 2016, Part I, LNCS 9916, pp. 1–10, 2016.
DOI: 10.1007/978-3-319-48890-5 1
2 H. Fan et al.
causes drift when object appearance varies. Babenko et al. [3] utilize the multiple
instance learning (MIL) method for visual tracking, which can alleviate drift to
some extent. Yang et al. [18] suggest a discriminative appearance model based
on superpixels, which facilitates the tracking algorithm to distinguish the target
from the background.
On the other hand, the generative models formulate tracking problem as
searching for regions most similar to object. These methods are based on either
subspace models or templates and update appearance model dynamically. In [16],
the incremental visual tracking method suggests an online approach for efficiently
learning and updating a low dimensional principal components analysis (PCA)
subspace representation for the object. However, this representation scheme is
sensitive to occlusion. Adam et al. [1] present a fragment-based template model
for visual tracking. Mei and Ling [15] model the object appearance with sparse
representation for visual tracking and achieve a good performance.
Though having achieved promising performance for object tracking, the
aforementioned algorithms often suffer from drifting problems when substantial
non-rigid and articulated motions are involved in the object.
To solve the problem of tracking non-rigid and/or articulated objects, we
propose a novel tracking algorithm with local superpixel matching and markov
random field (MRF). Our method mainly contains three stages. In the first stage,
we construct a superpixel dataset by segmenting training frames into superpix-
els, and each superpixel in the dataset is represented with multiple features.
Through this way, the appearance information of the object is encoded in the
superpixel dataset. In the second stage, for each new frame, we represent it with
its superpixels. We can compute its object-background confidence map by com-
paring its superpixels with their k-nearest neighbors in the superpixel dataset.
In this process, the tracking task is treated as separating object pixels from
background pixels. Taking the context information into consideration, we utilize
MRF to further improve the accuracy of the confidence map. In addition, the
local context information of each superpixel is incorporated through a feedback
to refine superpixel matching. In the final stage, object tracking is achieved via
searching the best candidate by maximum a posterior estimate based on the con-
fidence map. When tracking is completed in each frame, we collect good tracking
results to update the superpixel dataset. With the help of this update scheme,
our tracker is able to adapt to the appearance changes of the target. Figure 1
illustrates the framework of the proposed method.
For each new frame, we firstly extract the surrounding region1 of the target in the
last and then segment this region into superpixels with the same method in [11].
Let M be the number of its superpixels. For the ith superpixel sj (1≤ j ≤ M ),
we are able to calculate its label cost by comparing its k-nearest neighbor Nk (j)
in superpixel dataset D as follows
i∈N (j),yi =c K(xj , xi )
U (yj = c|sj ) = 1 − k (2)
i∈Nk (j) K(xj , xi )
where xj denotes the feature of sj , c ∈ {0, 1} represents the label and K(xj , xi )
is the intersection kernel between features xj and xi .
In this work, tracking is treated as separating object pixels from background
pixels. In order to exploit the context relationship of object pixels and back-
ground pixels, we utilize MRF inference for contextual constraints. The energy
function is given by
E(Y ) = U (yp = c) + λ V (yp = c, yq = c ) (3)
p pq
1
The surrounding region is a square area centered at the location of target Xtc , and
1
its side length is equal to λs [Xst ] 2 , where Xtc represents the center location of target
s
region Xt and Xt denotes its size. The parameter λs is a constant variable, which
determines the size of this surrounding region.
4 H. Fan et al.
where p, q are pixel indices, c, c are candidate labels and λ is the weight of
pairwise energy. The unary energy of one pixel is given by the superpixel it
belongs to
U (yp = c) = U (yj = c|sj ), p ∈ sj (4)
The pairwise energy on edges is given by spatially variant label cost
where d(p, q) = exp(−I(p) − I(q)2 /2σ 2 ) is the color dissimilarity between two
adjacent pixels, and μ(c, c ) is the penalty of assigning label c and c to two
adjacent pixels and defined by log-likelihood of label co-occurrence statistics
Through this way, we can derive the labels of all pixels by performing MAP
interference on E(Y ) with graph cut optimization in [4].
Taking into local context information of each superpixel into account, we
adopt a simple yet effective feedback mechanism as in [19]. In the feedback
process, we can obtain the pixel-wise classification likelihood of each pixel by
1
(p, c) = (7)
1 + exp(U (yp = c))
For robust superpixel matching, we exploit the local context of each super-
pixel. For superpixel sj , we divide its neighborhood into left, right, top, bottom
four cells {lc1j , lc2j , lc3j , lc4j } (see Fig. 2). For each cell lckj (1 ≤ k ≤ 4), we compute
its sparse context hkj = [hkj1 , hkj2 ] by
Through the above process, we are able to obtain the matching score Score(j)
for superpixel sj by
and the confidence map C for each pixel on the entire current frame as follows.
We assign every pixel whose label is object with 1, and every pixel whose label
is background or outside the surrounding region with −1. Figure 3 shows the
matching maps, confidence maps and tracking results of the target in video
Iceskater by our method.
Fig. 3. Matching maps, confidence maps and tracking results on video Iceskater. First
row: original images. Second row: matching maps of corresponding regions obtained by
our local superpixel matching. Third row: confidence maps of corresponding regions
derived by performing MRF on matching maps. Fourth row: the final tracking results
of each frame (see details in Sect. 2.3).
t by
represents the observation of target in frame τ , we can obtain estimation X
computing the maximum a posterior via
t = argmax p(X i |Y t )
X (10)
t
Xti
where X t denotes the ith sample at the state of Xt . The posterior probability
p(Xt |Z t ) can be obtained by the Bayesian theorem recursively via
i
p(Xt |Y t ) ∝ p(zt |Xt ) p(Xt |Xt−1 )p(Xt−1 |Z t−1 )dXt−1 (11)
where p(Xt |Xt−1 ) and p(zt |Xt ) represent the dynamic model and observation
model respectively.
The dynamic model indicates the temporal correlation of the target state
between consecutive frames. We apply affine transformation to model the target
motion between two consecutive frames within the particle filter framework. The
state transition can be formulated as
p(Xt |Xt−1 ) = N (Xt ; Xt−1 , Ψ ) (12)
where Ψ is a diagonal covariance matrix whose elements are the variance of
affine parameters. The observation model p(zt |Xt ) represents the probability of
the observation zt at state Xt . In this paper, the observation for ith sample at
the state of Xt is designed as in [18] by
p(zt |Xti ) ∝ vti (w, v) × [S(Xti )/S(Xt−1 )] (13)
(w,v)∈Cti
where Cti is the confidence map of the ith candidate warped from confidence
map of corresponding region, vti (w, v) denotes the confidence value of pixels at
location (w, v), S(Xti ) represents the area size of the ith candidate and S(Xt−1 )
is the area size of the object in last frame. Through Bayesian inference, we can
determine the candidate sample with the maximum observation as the tracking
result.
When heavy occlusion happens, the occlusion coefficient Ot will be large, and
thus it is unnecessary to add the tracking result into T . We set a threshold θ to
determine whether the tracking result is added into T . If Ot > θ, we skip this
frame to avoid introducing noise into T . Otherwise, we add the tracking result
into T and remove the oldest element from T if the number of elements in T is
larger than L.
3 Experiments
We evaluate our tracker on eight challenging image sequences and compare it
with seven state-of-the-art tracking methods. These algorithms are SPT tracking
[18], CT tracking [20], SCM tracking [23], STC tracking [21], ASLA tracking [12],
PCOM tracking [17], MTT tracking [22]. The proposed algorithm is implemented
Table 1. Average center location error (CLE) in pixel. The best and the second best
results are shown in red and blue fonts.
0 0 0 0
0 50 100 150 200 250 300 350 400 0 50 100 150 200 250 300 0 100 200 300 400 500 600 700 800 0 100 200 300 400 500
Frame Index Frame Index Frame Index Frame Index
100 80 80
100
60 60
50 40 40
50
20 20
0 0 0 0
0 5 10 15 20 25 0 20 40 60 80 100 120 140 160 0 50 100 150 200 250 300 350 400 450 0 20 40 60 80 100 120 140
Frame Index Frame Index Frame Index Frame Index
Fig. 4. Quantitative evaluation in terms of center location error in pixel. The proposed
method is compared with seven state-of-the-art algorithms on eight challenging test
sequences.
8 H. Fan et al.
in MATLAB and runs at 1.5 frames per second on a 3.2 GHz Intel E3-1225 v3
Core PC with 8 GB memory. The parameters of the proposed tracker are fixed in
all experiments. The number of neighbors k in Eq. (2) is set to 7. The number of
particles in Bayesian framework is 300 to 600. The λs is set to 1.5. The number
of initial training samples is 5. The length L of set T is fixed to 10, and H is set
to 5. The threshold θ is 0.8.
(a) Bikeshow
(b) David3
(c) Motocross2
(d) Transformer
ASLA CT MTT PCOM SCM STC SPT Ours
4 Conclusion
In this paper, we propose a novel method for object tracking, especially for the
targets involved with non-rigid and articulated motions. This approach mainly
consists of three stages. In the first stage, a superpixel database is constructed
to represent the appearance of object. In the second stage, when a new frame
arrives, it is firstly segmented into superpixels. Then we compute its confidence
via superpixel matching and MRF. Taking context information into account, we
utilize MRF to further improve the accuracy of confidence map. In addition,
the local context information is incorporated through a feedback to refine super-
pixel matching. In the last stage, visual tracking is achieved through finding the
best candidate by maximum a posterior estimate based on the confidence map.
Experiments evidence the effectiveness of our method.
References
1. Adam, A., Rivlin, E., Shimshoni, I.: Robust fragments based tracking using the
integral histogram. In: IEEE Conference on Computer Vision and Pattern Recog-
nition (CVPR), pp. 798–805 (2006)
2. Avidan, S.: Ensemble tracking. IEEE Trans. Pattern Anal. Mach. Intell. (TPAMI)
29(2), 261–271 (2007)
3. Babenko, B., Yang, M.-H., Belongie, S.: Robust object tracking with online mul-
tiple instance learning. IEEE Trans. Pattern Anal. Mach. Intell. (TPAMI) 33(8),
1619–1632 (2011)
4. Boykov, Y., Veksler, O., Zabih, R.: Fast approximate energy minimization via
graph cuts. IEEE Trans. Pattern Anal. Mach. Intell. (TPAMI) 23(11), 1222–1239
(2001)
Another random document with
no related content on Scribd:
(good)
Strawberry Isinglass Jelly 468
Fancy Jellies, and Jelly in 469
Belgrave mould
Queen Mab’s Pudding (an 470
elegant summer dish)
Nesselróde Cream 471
Crême à la Comtesse, or the 472
Countess’s Cream
An excellent Trifle 473
Swiss Cream, or Trifle (very 473
good)
Tipsy Cake, or Brandy Trifle 474
Chantilly Basket filled with 474
whipped Cream and fresh
Strawberries
Very good Lemon Cream, 475
made without Cream
Fruit Creams, and Italian 475
Creams
Very superior whipped 476
Syllabubs
Good common Blanc-mange, 476
or Blanc Manger (Author’s
receipt)
Richer Blanc-mange 477
Jaumange, or Jaune Manger; 477
sometimes called Dutch
Flummery
Extremely good Strawberry 477
Blanc-mange, or Bavarian
Cream
Quince Blanc-mange 478
(delicious)
Quince Blanc-mange, with 478
Almond Cream
Apricot Blanc-mange, or 479
Crême Parisienne
Currant Blanc-mange 479
Lemon Sponge, or Moulded 480
Lemon Cream
An Apple Hedgehog, or 480
Suédoise
Imperial Gooseberry-fool 480
Very good old-fashioned boiled 481
Custard
Rich boiled Custard 481
The Queen’s Custard 481
Currant Custard 482
Quince or Apple Custards 482
The Duke’s Custard 482
Chocolate Custards 483
Common baked Custard 483
A finer baked Custard 483
French Custards or Creams 484
German Puffs 484
A Meringue of Rhubarb, or 485
green Gooseberries
Creamed Spring Fruit, or 486
Rhubarb Trifle
Meringue of Pears, or other 486
fruit
An Apple Charlotte, or 486
Charlotte de Pommes
Marmalade for the Charlotte 487
A Charlotte à la Parisienne 486
A Gertrude à la Créme 486
Pommes au Beurre (Buttered 488
Apples) (excellent)
Suédoise of Peaches 488
Aroce Doce, or Sweet Rice à la 489
Portugaise
Cocoa Nut Doce 490
Buttered Cherries (Cerises au 490
Beurre)
Sweet Macaroni 490
Bermuda Witches 491
Nesselróde Pudding 491
Stewed Figs (a very nice 492
Compote)
CHAPTER XXIV.
PRESERVES.
Page
PICKLES.
Page
CAKES.
Page
CONFECTIONARY.
Page
DESSERT DISHES.
Page
Page
Page
Coffee 587
To roast Coffee 588
A few general directions for 589
making Coffee
Excellent Breakfast Coffee 590
To boil Coffee 591
Café Noir 592
Burnt Coffee, or Coffee à la 592
militaire (In France vulgarly
called Gloria)
To make Chocolate 592
A Spanish recipe for making 592
and serving Chocolate
To make Cocoa 593
CHAPTER XXXI.
BREAD.
Page
Page