A Unified Shot Boundary Detection Framework Based on Graph Partition Model Jinhui Yuan



Jianmin Li

[email protected]

Bo Zhang

{lijianmin,linfz,dcszb}@mail.tsinghua.edu.cn

ABSTRACT

which shows superiority to the existing methods when evaluated on the platform of TRECVID benchmark. In the previous work [4], we propose a graph partition model with temporal constraints to perform temporal data segmentation. The model tries to minimize the association between the two subgraphs while maximize the association within each subgraph with the following criterion:

In this paper, we propose a unified shot boundary detection framework by extending the previous work of graph partition model with temporal constraints. To detect both the abrupt transitions (CUTs) and gradual transitions (GTs, excluding fade out/in) in a unified way, we incorporate temporal multi-resolution analysis into the model. Furthermore, instead of ad-hoc thresholding scheme, we construct a novel kind of feature to characterize shot transitions and employ support vector machine (SVM) with active leaning strategy to classify boundaries and non-boundaries. Extensive experiments have been carried out on the platform of TRECVID benchmark. The experimental results show that the proposed framework outperforms some others and achieves satisfactory results.

M cut(A, B) =

cut(A, B) cut(A, B) + . assoc(A) assoc(B)

(1)

For temporal data, the optimal segmentation can be got at some i with the minimal score (i ∈ {1, · · · , N − 1}): score(i) = M cut({1, 2, · · · , i}, {i + 1, i + 2, · · · , N }). (2) In this paper we apply and extend the previous work to fulfil temporal video segmentation. We incorporate a novel temporal multi-resolution analysis algorithm into the model thus the CUTs and GTs can be detected in a unified way. In addition, with a novel constructed feature, we employ SVM with active learning strategy to classify boundaries and nonboundaries. The remainder of the paper is organized as follows. Section 2 introduces how to employ the graph partition model to perform temporal video segmentation. Section 3 presents the temporal multi-resolution method. Section 4 introduces the support vector machine framework with active learning strategy. Section 5 evaluates the algorithms on the platform of TRECVID benchmark. And finally we conclude the paper in Section 6.

Categories and Subject Descriptors H.3 [Information Storage and Retrieval]: Content Analysis and Indexing—abstracting methods

General Terms algorithms, management

Keywords graph partition, temporal multi-resolution, active learning

1.

Fuzong Lin

Department of Computer Science and Technology, Tsinghua University Beijing, 100084, P. R. China

Department of Computer Science and Technology, Tsinghua University Beijing, 100084, P. R. China

INTRODUCTION 2. TEMPORAL VIDEO SEGMENTATION WITH GRAPH PARTITION MODEL

Shot boundary detection is a prerequisite step of content based video retrieval. A large number of methods have been proposed to perform shot boundary detection. In this paper, we propose a novel shot boundary detection framework

Given a video sequence, we treat each frame as a node and link each other with an edge, and a weighted graph G(V, E) can be constructed. Thus, the shot boundary detection is formulated as a graph partition problem.

∗Supported by National Natural Science Foundation of China (60135010, 60321002) and Chinese National Key Foundation Research & Development Plan (2004CB318108).

2.1

How to define the edge weight wij ?

Let H i be a k bins color histogram of the i-th frame and adopt histogram intersection method to measure the similarity, a reasonable definition of wij is: ( −ki−jk2 2 X min(Hki , H j ) k e σ2 if |i − j| < r2 (3) wij = × i H 0 otherwise. k k

Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. To copy otherwise, to republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. MM’05, November 6–11, 2005, Singapore. Copyright 2005 ACM 1-59593-044-2/05/0011 ...$5.00.

where σ is a factor reflecting the similarity decaying with the temporal interval increasing, and r denotes the maxi-

539

1

Active Matrix

10

10

20

20

30

30

40

40

50

50

60

60

2

3

10

4

20

5

6

7

8

30

9

10 70

70 10

20

30

40

50

60

70

10

20

30

40

50

60

70

1

2

3

4

5

6

7

8

9

10

40

Figure 2: The three matrices are Wδi , Mδ , Vri respectively, where r = 10 and δ = 7. The operation “¯” means the Hadamard multiplication. The white blocks in Mδ represent entries 1. The first two matrices are square ones of width 70, while the last one is square matrix of width 10.

50

60

Fa lshg il ht

10

20

(a)

30

40

50

60

(b)

1

1

0.95

0.9

0.9

Flashlight

0.8

Score of min−max cut

Histogram Intersection

0.85

0.75

0.7

Flashlight

based on pair-wise comparison, the three corresponding sharp valleys are usually considered as boundaries of shots. Consequently, many false alarms are caused. While in the proposed approach, the strong connectivity among the frames before and after the flashlight frame makes it unlikely to separate the sequence to two parts. The method is robust to various abrupt illumination change, and thus no specific flashlight detector is needed. The experiment in Section 5 will further confirm this analysis.

0.8

0.7

0.65

0.6

0.6

0.55

0.5

0.5 0

10

20

30 40 Frame Index

50

60

70

10

(c) bbbbb

20

30 40 Frame Index

50

60

(d)

Figure 1: (a): A video sequence with three flashlights occurring. (b): The visualization of the corresponding similarity matrix, in which the red rectangle indicates the range of active matrix. (c): The curve obtained by the comparison between the successive frames. (d): The curve obtained according to Equation 4.

3. TEMPORAL MULTI-RESOLUTION ANALYSIS By observing the patterns on the similarity matrix, we can find that there is a clear “chessboard” pattern for CUT boundary, while for GT boundary, it may yield a blurry pattern on the similarity matrix. Conversely, noticing that, at a lower resolution, whatever the length of the GT is, there will always be a “chessboard” pattern clear enough. To detect CUTs and GTs simultaneously, we re-define the score of the i-th feasible cut as follows: r r score(i, δ) = M cut({i− ×δ, · · · , i−δ, i}, {i+δ, · · · , i+ ×δ}). 2 2 (5) where δ is the sampling rate of the frames, δ ∈ {1, 2, · · · }. Equation 5 means that when calculating the score of the i-th feasible cut, instead of involving all the frames in a neighborhood of {i−r2 , · · · , i+r2 }, the algorithm only samples every δ frames in a larger range of {i− r2 ×δ, · · · , i+ r2 ×δ}. With the δ varying, multiple temporal resolution graphs can be constructed. To facilitate the computation, we define a square selective matrix Mδ of width r × δ, in which entry 0 indicating the the corresponding frame is not sampled, 1 representing the frame sampled. Let Wδi be the square sub-matrix of W, centering at i and with the width r×δ. To calculate score(i) at the resolution δ, the algorithm just performs Hadamard multiplication of Wδi and Mδ , and thus results in a new matrix Vδi , in which the entries corresponding to the nonsampled frames equal 0. Furthermore, ignoring the 0 entries, Vδi can be restricted an equivalent but smaller square matrix Vri of width r. The above process is depicted in Figure 2.

mum range in which the frames are considered to influence each other. Therefore, the calculation is restricted in a r×r sub-matrix, which we call active matrix. Consequently, the Equation 2 can be simplified as: r r (4) score(i) = M cut({i− , · · · , i}, {i+1, · · · , i+ }). 2 2

2.2

The Algorithm

In summary, the video temporal segmentation algorithm consists of the following steps:

Step 1. Given a video sequence, treat each frame as a node and link each other by an edge, to construct a weighted graph G(V, E). Step 2. Compute wij according to Equation 3, obtaining a similarity matrix W. Step 3. Calculate scores of the N−1 feasible cuts according to Equation 4. Step 4. Select feasible cuts whose scores are the local minima of the corresponding neighborhoods within a radius of r2 . Step 5. Declare the cuts whose scores are below a predefined threshold as CUTs.

2.3

Analysis of the Algorithm

4. SUPPORT VECTOR MACHINE ACTIVE LEARNING

One prominent advantage of the approach is the robustness. Instead of pair-wise comparison by traditional novelty detection, the graph partition method performs boundary detection by considering the feature variation in a local neighborhood. As shown in Figure 1, there are three flashlights occurring during the video sequence. In the method

On the curve of scores, each boundary corresponds to a sufficiently small local minimum. However, not every local minimum is a shot boundary. Only by evaluating the magnitudes of the local minima can not successfully distin-

540

guish boundaries and non-boundaries. In this paper, we propose to employ SVM with active learning strategy to classify boundaries and non-boundaries according to the shapes of local minima.

4.1

1

0.95

0.9

Feature Construction

0.85

precision

Formally, let Siδ denote score(i, δ), defined by Equation 5. Then the feature, which characterizes the shape of the valley centering at i on the δ resolution curve, is define as:

0.75

δ δ δ δ δ r ×δ , Si−( r −1)×δ , · · · , Si , · · · , Si+( r −1)×δ , Si+ r ×δ ] fiδ = [Si− 2 2 2 2 (6) Obviously, for each δ, the fδ is a (r +1) dimension feature. To make full use of the information across different resolutions, i.e. δ ∈ {1, 3, 5, 7, 9}, we concatenate those features to construct a new feature vector as:

Fi =

[fi1 , fi3 , fi5 , fi7 , fi9 ]

0.7

0.65

0.6 0.82

0.84

0.86

0.88

0.9

0.92

0.94

0.96

0.98

1

(7) Figure 3: Performance comparison with other related algorithms.

4.2 Active Learning Strategy

5.2 Performance Comparison

Manually labeling the local minima as boundaries or nonboundaries to obtain a training set is a tedious job. Similar to the idea of [3], one heuristic active learning criterion is employed. We assume that the examples difficult for thresholding method to clearly classify lie near the dividing hyperplane of the SVM. We firstly collect all the valleys which are under a specified threshold θ from all the available local minima. These valleys collected, including real boundaries and various false alarms, constitute the training set. Then the real boundaries are labeled as positive examples, and the false alarms are labeled as negative examples. Here the threshold θ is set low enough to guarantee almost all of the boundaries are collected. Meanwhile, with the threshold θ, we can remove a lot of local minima which will not influence the position of the SVM’s hyperplane.

5.2.1 Graph Partition Model To show the effectiveness of the graph partition model, we implement five other related approaches and evaluate them on the “DATA NO GT”. They are:

Thresholding Directly compare the difference of successive frames. Min-max cut The algorithm proposed in Section 2.2. Scale space Kernel correlation depicted in [2]. Diagonal CS Kernel correlation depicted in [2]. Cross S Kernel correlation depicted in [2]. Full S Kernel correlation depicted in [2].

EXPERIMENTS

As the Figure 3 shows, all the five approaches via multi-pair comparison outperform the pair-wise comparison method “Thresholding”. Both the “Diagonal CS” and the “Cross S” kernels emphasize the dissimilarity between the different shots, and they performs almost the same. The “Full S” somehow outperforms the proposed “Min-max cut” method and both of them perform best. This is not surprising, since they both consider the similarity between different shots and within the the same shot. In fact, the “Full S” is equivalent to an alternative definition of min-max cut:

In this section, we evaluate the proposed framework on the platform of TRECVID benchmark [1]. Several comparison experiments have been designed to justify the ideas of this paper.

5.1

Thresholding Min−max cut Scale space Diagonal CS Cross S Full S

recall

Given r = 10, Fi is a 55 dimension feature including the the i-th feasible cut’s shape information at different resolutions.

5.

0.8

Experimental Setup

All the 2003 and 2004 TRECVID test collections for the task of shot boundary detection (SBD) are adopted. So far we have not considered incorporating the fade out/in (FOI) detection in the framework. Therefore, we re-edit the video collections by transforming FOIs into CUTs. We call the video collection without FOIs as “DATA NO FOI”. On the other hand, in order to focus on the comparison of the characteristics of different algorithms, we apply the algorithms to fulfil CUT detection instead of boundary detection in several experiments. Thus, we create another video collection without GTs by re-editing the GTs into CUTs. This collection is called as “DATA NO GT”. The algorithms evaluated on the this data set are all single resolution implementations, i.e. δ = 1. Similar to other information retrieval task, the performance is evaluated by recall and precision criteria. To rank performance of different algorithms, F1 measure, a harmonic average of recall and precision is adopted.

M cut(A, B) = assoca(A)+assoca(B)−2×cut(A, B) (8) Note that our experimental result is inconsistent with that of Cooper [2], in which the author claims that the “Full S” performs worst. However, we believe that our results are more reliable. Firstly, we evaulate them on a more delicate data set of “DATA NO GT”. Secondly, we design a more straightforward experimental setup, in which a simple histogram feature and thresholding method are adopted, while Cooper has employed multi-scale features and KNN to classify CUTs and non-CUTs.

5.2.2 Temporal Multi-resolution Analysis To examine the effectiveness of temporal multi-resolution analysis, we implemented four different methods and evalated

541

1

1

0.95 0.98 0.9 0.96

F1−measure

precision

0.85

0.8

0.94

0.75 0.92 0.7

0.65

0.6 0.8

0.82

Graph+Threshold Graph+SVM Graph+ASVM

0.9

delta=1 delta=5 delta=7 delta=1 & 5 0.84

0.86

0.88

0.9

0.92

0.94

0.96

0.88

0.98

1

2

3

4

5

recall

6

7

8

9

10

11

12

Video index

Figure 4: Evaluation of temporal multi-resolution analysis.

Figure 5: Support vector machine with active learning strategy.

them on “DATA NO FOI”:

As shown in Figure 5, both the SVM methods have outperformed the thresholding method. The performance of “Graph+ASVM” is comparable to that of “Graph+SVM”, while the size of the training set for active learning is just about 16 of that of “Graph+SVM”, 1090 for the former and 6948 for the latter.

delta=1 delta=5 delta=7 delta=1 & 5

δ = 1 for the algorithm of Section 2.2. δ = 5 for the algorithm of Section 2.2. δ = 7 for the algorithm of Section 2.2. Multi-resolution analysis, combining the δ = 1 and δ = 5 for the algorithm of Section 2.2.

6.

CONCLUSIONS AND FUTURE WORK

In this paper, we extend our previous work and design a unified shot boundary detection framework based on graph partition model. Extensive experiments on the benchmark of TRECVID have been performed to justify the proposed ideas. Nevertheless, several problems remains for further research. For example, we have not incorporated the detection of FOIs into the framework. With multi-resolution analysis, we have to design methods to effectively reduce the disturbances of motion. How to effectively make use of information across different resolutions is also an important problem to solve. Despite the remaining problems above, we believe that the proposed framework may be a promising method for nearly perfect temporal segmentation of videos.

As shown in Figure 4, delta=5 outperforms delta=1. That is because at high resolution, most of the long GTs are missed. However, delta=7, which is the lowest resolution, performs worse than the other two higher resolution methods. Then why the lowest resolution performs worst? We have examined the detection result of delta=7 and find that, although the lower resolution helps to detect long GTs, it also suffers the enlargement of disturbances of motion. Since we have not incorporate the post processing with motion filtering, the delta=7 performs worst. Finally, the delta=1 & 5, with the multi-resolution analysis, performs best. We can expect that, if we import the analysis of motion activity, the performance of the proposed method will be further improved.

7. REFERENCES [1] Trec video retrieval evaluation, http://www-nlpir.nist.gov/projects/trecvid/. [2] M. Cooper. Video segmentation combining similarity analysis and classification. In ACM Multimedia, October 2004. [3] G. Schohn and D. Cohn. Less is more: Active learning with support vector machines. In Proc. 17th International Conf. on Machine Learning, pages 839–846. Morgan Kaufmann, San Francisco, CA, 2000. [4] J. Yuan, B. Zhang, and F. Lin. Graph partition model for robust temporal data segmentation. In Proc. of PAKDD, LNAI, volume 3518, pages 758–763, 2005.

5.2.3 SVM with Active Learning Strategy The experiments are performed on “DATA NO GT” and a single resolution feature of δ ∈ {1} is adopted. Two of the 2003 test collections are chosen for training and all the twelve videos of 2004 test collections are used for testing. RBF(radial basis function) is adopted as the kernel function. C = 400 and σ = 1.0 are selected after a cross validation process. For the “Graph+Threshold” method, the best result among various threshold settings is chosen to be compared. Graph+Threshold Algorithm of Section 2.2. Graph+SVM SVM Trained with randomly selected samples. Graph+ASVM SVM Trained with active learning strategy.

542

A Unified Shot Boundary Detection Framework ... - Semantic Scholar

Nov 11, 2005 - Department of Computer Science and. Technology, Tsinghua University ..... the best result among various threshold settings is chosen to.

1MB Sizes 22 Downloads 290 Views

Recommend Documents

A novel shot boundary detection framework - Semantic Scholar
Fuzong Lin and Bo Zhang. State Key Laboratory of Intelligent Technology and System. Department of Computer Science and Technology. Tsinghua University ...

Linear Network Codes: A Unified Framework for ... - Semantic Scholar
This work was supported in part by NSF grant CCR-0220039, a grant from the Lee Center for. Advanced Networking, Hewlett-Packard 008542-008, and University of ..... While we call the resulting code a joint source-channel code for historical ...

A Unified Framework and Algorithm for Channel ... - Semantic Scholar
with frequency hopping signalling," Proceedings of the IEEE, vol 75, No. ... 38] T. Nishizeki and N. Chiba, \"Planar Graphs : Theory and Algorithms (Annals of ...

Linear Network Codes: A Unified Framework for ... - Semantic Scholar
Page 1 ..... For any n × ⌊nR⌋ matrix bn, we can build a linear channel code with .... For any n × n matrix cn, we can build a joint source-channel code for the.

A Unified Framework of HMM Adaptation with Joint ... - Semantic Scholar
that the system becomes better matched to the distorted environment. ...... Incremental online feature space MLLR adaptation for telephony speech recognition.

A Unified Framework of HMM Adaptation with Joint ... - Semantic Scholar
used as a simple technique to reduce additive noise in the spectral domain. ... feature space rotation and vocal tract length normalization to get satisfactory ...

A Formal Study of Shot Boundary Detection - CiteSeerX
Color versions of one or more of the figures in this paper are available online at http://ieeexplore.ieee.org. ... not obtained by thresholding schemes but by machine learning ...... working toward the M.S. degree in the Department of Computer ...

A Formal Study of Shot Boundary Detection - CiteSeerX
Based on the comparison of the existing approaches, optimal criteria for ... in computer perfor- mance and the growth of the Internet, have led to the widespread.

A Formal Study of Shot Boundary Detection
fied shot boundary detection system based on graph partition ... technologies and tools for efficient indexing, browsing and re- ...... Right: Visualization of the similarity matrix of the left graph. w is defined as the reciprocal of Euclidean dista

Toward a Unified Artificial Intelligence - Semantic Scholar
Department of Computer and Information Sciences. Temple University .... In the following, such a system is introduced. ..... rewritten as one of the following inheritance statements: ..... goal is a matter of degree, partially indicated by the utilit

Toward a Unified Artificial Intelligence - Semantic Scholar
flexible, and can use the other techniques as tools to solve concrete problems. ..... value in an idealized situation, so as to provide a foundation for the inference.

Unified Inference in Extended Syllogism - Semantic Scholar
duction/abduction/induction triad is defined formally in terms of the position of the ... the terminology introduced by Flach and Kakas, this volume), cor- respond to ...

Boundary Element Formulation of Harmonic ... - Semantic Scholar
On a deeper level, BEM makes possible the comparison of trans- finite harmonic ... Solving a Dirichlet problem could seem a high price to pay, but the quality of the .... Euclidean space, and not just to some large yet bounded domain. This task ...

Boundary Element Formulation of Harmonic ... - Semantic Scholar
that this will allow tailoring new interpolates for particular needs. Another direction for research is ... ment methods with the software library BEMLIB. Chapman &.

Unified Inference in Extended Syllogism - Semantic Scholar
... formally in terms of the position of the shared term: c© 1998 Kluwer Academic Publishers. ...... Prior Analytics. Hackett Publishing Company, Indianapolis, Indi-.

a multimodal search engine based on rich unified ... - Semantic Scholar
Apr 16, 2012 - Copyright is held by the International World Wide Web Conference Com- ..... [1] Apple iPhone 4S – Ask Siri to help you get things done. Avail. at.

21. UniGene: A Unified View of the Transcriptome - Semantic Scholar
Oct 9, 2002 - introduced the term EST to refer to this new class of sequence, which is .... At present, UniGene focuses on protein-coding genes of the nuclear genome ... query bar and a sidebar providing links to related online resources.

a multimodal search engine based on rich unified ... - Semantic Scholar
Apr 16, 2012 - Google's Voice Actions [2] for Android, and through Voice. Search [3] for .... mented with the objective of sharing one common code base.

Optimal Detection of Heterogeneous and ... - Semantic Scholar
Oct 28, 2010 - where ¯Φ = 1 − Φ is the survival function of N(0,1). Second, sort the .... (β;σ) is a function of β and ...... When σ ≥ 1, the exponent is a convex.

Intrusion Detection Visualization and Software ... - Semantic Scholar
fake program downloads, worms, application of software vulnerabilities, web bugs, etc. 3. .... Accounting. Process. Accounting ..... e.g., to management. Thus, in a ...

Intrusion Detection Visualization and Software ... - Semantic Scholar
fake program downloads, worms, application of software vulnerabilities, web bugs, etc. 3. .... Accounting. Process. Accounting ..... e.g., to management. Thus, in a ...