ArticlePDF Available

The MPEG7 Video Signature Tools for Content Identification

Authors:

Abstract

This paper presents the core technologies of the video signature tools recently standardized by ISO/IEC Moving Picture Experts Group (MPEG) as an amendment to the MPEG-7 Standard (ISO/IEC 15938). The video signature is a high-performance content fingerprint that is suitable for desktop scale to web-scale deployment and provides high levels of robustness to common video editing operations and high temporal localization accuracy at extremely low false alarm rates, achieving a detection rate in the order of 96% at a false alarm rate in the order of five false matches per million comparisons. The applications of the video signature are numerous and include rights management and monetization, distribution management, usage monitoring, metadata association, and corporate or personal database management. In this paper, we review the prior work in the field, explain the standardization process and status, and provide details and evaluation results for the video signature tools.
IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS FOR VIDEO TECHNOLOGY TCSVT 5350
1
Abstract—This paper presents the core technologies of the
Video Signature Tools recently standardized by ISO/IEC MPEG
as an amendment to the MPEG-7 Standard (ISO/IEC 15938).
The Video Signature is a high-performance content fingerprint
which is suitable for desktop-scale to web-scale deployment and
provides high levels of robustness to common video editing
operations and high temporal localization accuracy at extremely
low false alarm rates, achieving a detection rate in the order of
96% at a false alarm rate in the order of five false matches per
million comparisons. The applications of the Video Signature are
numerous and include rights management and monetization,
distribution management, usage monitoring, metadata
association, and corporate or personal database management. In
this paper we review the prior work in the field, explain the
standardization process and status, and provide details and
evaluation results for the Video Signature Tools.
Index Terms—content description, content identification,
content localization, content fingerprint, MPEG, video
fingerprint, video signature, video identification
I. INTRODUCTION
HE amount of video content which people generate and
consume has been increasing at a stunning pace in recent
years. According to figures released by the company itself, in
2010 two billion videos a day were watched on YouTube and
hundreds of thousands of videos were uploaded daily, at a rate
of 24 hours of content every minute [1]. Instantaneous access
S. Paschalakis is with Visual Atoms Ltd, The Surrey Technology Centre,
40 Occam Road, Guildford Surrey GU2 7YG, UK (tel: +44 (0)1483 688181,
email: s.paschalakis@visualatoms.com). He was with Mitsubishi Electric
R&D Centre Europe (MERCE).
K. Iwamoto is with NEC Corporation, 1753, Shimonumabe, Nakahara-Ku,
Kawasaki, Kanagawa 211-8666, Japan (email: k-iwamoto@ay.jp.nec.com)
P. Brasnett is with Imagination Technologies Ltd, Imagination House,
Home Park Estate, Kings Langley, Hertfordshire, WD4 8LZ, UK (email:
paul.brasnett@imgtec.com). He was with MERCE.
N. Sprljan is with Maxeler (email: nsprljan@maxeler.com). He was with
Visual Atoms Ltd and MERCE.
R. Oami is with NEC Corporation (email: r-oami@az.jp.nec.com)
T. Nomura is with NEC Corporation (email: t-nomura@da.jp.nec.com)
A. Yamada is with NEC Corporation (email: a-yamada@da.jp.nec.com)
M. Bober is with The University of Surrey and Visual Atoms Ltd (email:
m.bober@visualatoms.com). He was with MERCE.
Copyright (c) 2011 IEEE. Personal use of this material is permitted.
However, permission to use this material for any other purposes must be
obtained from the IEEE by sending an email to
pubs-permissions@ieee.org
.
to the vast volumes of multimedia content available on the
web is something that we take for granted thanks to text-based
search engines. Despite this, there are no standard tools which
one can use to efficiently search for a copy of a specific piece
of video content, possibly in an edited or modified version and
embedded in a longer piece of video content, either on the web
or in one’s own personal database.
In recent years, video identification has been studied mainly
in two different, but complementary, ways. Firstly, as a
general problem of designing unique and robust fingerprints to
allow the identification and localization of a video embedded
in an unrelated longer video [2],[3]. Secondly, as the more
specific problem of near-duplicate video clip detection in
large databases [4],[5]. The latter investigations have focused
on the design of complete retrieval systems, including
fingerprint extraction and database indexing schemes, for the
retrieval of short near-duplicate clips exhibiting a complete or
significant temporal overlap. Such systems are not, by and
large, designed for the identification and localization of a
video embedded in a longer video, although there are
exceptions [6]. This paper focuses on the first problem of
designing video fingerprinting tools that achieve high levels of
robustness to common video editing operations and are able to
accurately detect and localize a piece of video content
embedded in a longer piece of unrelated video content, with
exceptionally low false alarm rates, in the order of only a few
false alarms per million comparisons.
To attain an interoperable solution for video content
identification, the Moving Picture Experts Group (MPEG), a
working group of ISO/IEC, issued a worldwide open call for
proposals in 2008 [7],[8]. The Video Signature, or Video
Signature Tools to use its full name, has now been
standardized and is the latest amendment to the MPEG-7
Standard, also known as ISO/IEC 15938 Multimedia Content
Description Interface [9],[10].
The rest of this paper is organized as follows. Section II
looks at related works in the field of interest. Section III
presents the development process of the Standard, while
Section IV looks at application scenarios for the Video
Signature. Sections V and VI examine the Video Signature
extraction and compression respectively. The matching and
localization process used in the development of the Standard
is presented in Section VII. The MPEG-7 evaluation
methodology and results are presented in Section VIII, and a
The MPEG-7 Video Signature Tools for Content
Identification
S. Paschalakis, Member, IEEE, K. Iwamoto, P. Brasnett, N. Sprljan, R. Oami, Member, IEEE,
T. Nomura, Member, IEEE, A. Yamada and M. Bober, Member, IEEE
T
Accepted version of IEEE-copyrighted article. (C) 2012 IEEE
Digital Object Identifier: 10.1109/TCSVT.2012.2189791
IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS FOR VIDEO TECHNOLOGY TCSVT 5350
2
brief comparative discussion based on recently reported
results in video identification is presented in Section IX.
Finally, Section X presents our concluding remarks.
II. R
ELATED WORKS
This section provides an overview of the key design choices
in creating a video fingerprint and then examines a number of
notable studies in the field, both for identification and
localization of a video embedded in unrelated video content
and for near-duplicate video clip detection.
Most video signatures rely on frame-level features. In turn,
such features are usually keypoint-based [11],[12],[13],[14],
block-based [2],[15],[16] or global [17],[18]. For keypoint-
based approaches, SIFT features [14],[19],[20] or local
descriptors around Harris interest points [12],[13],[21],[22]
are a common design choice. Such approaches can attain
increased robustness to content modifications, but also entail
increased computational costs. Block-based approaches
[2],[15],[16] involve the calculation of certain properties of
pre-defined spatial regions in the frame. Block-based methods
typically entail reduced computational costs, but are also less
robust to transformations such as scaling and rotation. Finally,
global features, such as the color histograms [17],[18] are the
most computationally efficient, but typically have less
discriminating power than keypoint-based and block-based
descriptions and poor tolerance to global modifications, such
as global color changes.
Another key design choice is whether a video signature will
include temporal features or not. Spatial video signatures
[2],[14] are a common design choice and are typically
structured as a sequence of frame signatures. The underlying
temporal relation between the frame signatures is then used
only during the matching or retrieval process. In contrast to
spatial video signatures, temporal video signatures actually
contain temporal features, e.g. based on motion statistics [13],
or by extraction of features jointly from groups of frames
[12],[15]. Clearly, a video signature may have both a spatial
component and a temporal component, which is the case with
the MPEG-7 Video Signature, as will be seen later on.
Furthermore, a video signature may be viewed as
temporally dense or sparse. Here, we use the term dense to
refer to a video signature which describes every frame in the
video, as in [2],[13]. Conversely, sparse video signatures
operate at the keyframe level, with keyframes typically
extracted at the shot level [11] or at fixed intervals [18].
Although the above can be viewed as the primary design
choices, there are other choices which will have an impact on
identification performance and computational efficiency, such
as extraction from native resolution or sub-sampled frames,
the use of intensity or color information, etc. All these factors
need to be correctly balanced in the signature design. For
example, keypoint-based frame signatures are typically
employed in sparse video signatures [12] to ensure that their
extraction and matching are computationally tractable.
Hampapur et al. studied the problem of identification of
video content embedded in longer video content in [2]. That
early evaluation used only a few hundred copy queries 1.3 to
20.3 seconds in length, subjected to only a single modification
from the original, namely a resolution change. In that
evaluation, a block-based ordinal signature achieved the
highest performance, with nearly 0% false positive and false
negative rates for 10.3 second clips. The signature was
calculated at each frame by dividing it into blocks, calculating
the mean intensity of each block, sorting them in ascending
order and then using the rank vector as the frame feature.
However, for shorter queries, the best performance reported in
[2] drops to only ~94% recall and precision. Furthermore, [2]
does not report the matching segment localization accuracy.
A more recent evaluation of video signatures was carried
out by Law-To et al. [3]. That evaluation was performed under
two different scenarios. The first scenario used 72 5-second
copy segments embedded in unrelated content and subjected
to a single modification chosen from contrast change,
cropping, blurring, letter-box, logo overlay or zoom. That
experiment showed a block-based temporal ordinal signature,
based on the mean intensities of frame blocks ranked in the
temporal direction, achieving the highest performance, with
perfect recall and precision. The second scenario used 60 copy
segments between one and 10 seconds embedded in unrelated
content and subjected to a random combination of
modifications. In that experiment a keypoint-based signature
achieved the highest identification performance, with 82%
recall at 95% precision. That signature was based on keypoint
temporal trajectory properties, calculated from a differential
description of the region around Harris interest points
identified in keyframes [13]. However, Law-To et al. note that
their evaluation was limited by the size of the database, which
contained only three hours of video and few queries.
TRECVid [23],[24] has also been conducting content-based
copy detection evaluations. TRECVid 2009 completed in
March 2010 [25] and included video-only experiments,
conducted on 938 copy queries and 469 non-copy queries,
with a total duration of ~7.3 hours, exhibiting modifications
including picture-in-picture, pattern insertion, re-encoding,
frame dropping, cropping, closed captioning, etc. The primary
performance measure used was minNDCR (minimum
Normalized Detection Cost Rate) for two different profiles,
“NOFA” (No False Alarm) and “Balanced”. minNDCR does
not measure the copy localization accuracy, which was
reported as secondary information. One of the best performing
methods in TRECVid 2009 was proposed by Liu et al. [11].
That method utilizes sparse temporal sampling, representing
each shot by a single keyframe, describes the frame contents
using SIFT features, and uses locality sensitive hashing (LSH)
for efficient indexing and query. Copy detection is based on a
gradual keyframe-level query process, followed by video-level
merging. The method achieved an average minNDCR of ~0.37
and ~0.38 for the NOFA and Balanced profiles respectively.
For NOFA, this translates to a detection rate of ~63%. For
Balanced, it is more difficult to determine the actual success
IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS FOR VIDEO TECHNOLOGY TCSVT 5350
3
and false alarm rates. An average localization accuracy of
~85% was achieved for both profiles, with a mean processing
time per query between 100 and 200 seconds.
In [4] a near-duplicate video clip detection system is
presented, comprising two detection schemes. The first
scheme uses the bounded coordinate system (BCS), a compact
representation which ignores temporal information and
summarizes each entire video via a single vector. The second
scheme utilizes frame symbolization (FRAS), mapping each
video to a sequence of symbols and taking temporal order into
account. Database indexing is used to reduce the search space
and improve matching speeds. The system was tested on
~11,000 TV commercials, each with duration of ~60 seconds.
The results show BCS achieving the best performance, with
~60% recall for ~90% precision and an average search time of
~50ms. However, the ROC curves in [4] show that the recall
rate drops to ~20% for a precision of ~95%, indicating that the
system would not achieve high recall rates at very high
precision levels. Since the system of [4] is designed to match
entire short clips, it does not address the issue of detecting and
localizing a video segment inside a longer unrelated video.
In [5], another near-duplicate video clip detection system is
presented, where videos are represented by a compact
spatiotemporal feature relying on relative grey-level intensity
distributions within a frame and the temporal structure along
the frame sequence. For retrieval, an inverted file index
structure is used to achieve high speeds. The system was
evaluated on a database of 12,790 videos retrieved from the
web using 24 different queries. A mean average precision
(MAP) of 0.953 was achieved, with an average query speed of
3.7ms. To demonstrate the system’s scalability, it was also
evaluated on an extended database of ~50,000 web videos,
TABLE
I
S
UMMARY OF
S
ELECTED
V
IDEO
I
DENTIFICATION
M
ETHODS AND
E
XPERIMENTS
P
ART
I
Method
Detecti
on
Type
Localization
Type
Data Modifications
MPEG-7
Video Sig.
Embedded
segments
Fully integrated:
accurate to 1sec
Independence: ~70,000 30-sec queries to 1,900 180-sec originals
Robustness: ~70,000 30-sec queries to 545 180-sec originals
Query matching segment length: 2 sec, 5 sec, 10 sec
Wide range manually
applied and tested
Liu et al.
[11]
Embedded
segments
Not integrated:
diagnostic only
~7.3 hours DB
938 copy queries, 469 non-copy queries, average length 32 sec
Wide range manually
applied and tested
Hampapur
et al. [2]
Embedded
segments
Not reported
~2 hours DB
Few hundred queries, 1.3 sec to 20.3 sec
Resolution reduction
Law-To
et al. [3]
Embedded
segments
Not reported
~3 hours DB
(a) 9 15-min queries, each with 8 5-sec copy segments
(b) 1 30-min query with 60 copy segments between 1 sec and 10 sec
Wide range manually
applied and tested
Paisitktiangk
rai et al. [6]
Embedded
segments
(coarse)
Not reported
12,790 web videos (24 text queries)
24 queries used in experiments
Those already
naturally present in
content
Shen et al.
[4]
Whole
clips
NA
~11,000 TV ads at ~60 sec each
20 query clips
Those already
naturally present in
content
Shang et al.
[5]
Whole
clips
NA
(a) 12,790 web videos (24 text queries)
(b) ~50,000 web videos including (a)
24 queries
Those already
naturally present in
content
Sarkar et al.
[18]
Whole
clips
NA
~38,000 web videos main DB
~21,600 query videos with modifications (from 1,200 videos of DB)
Wide range manually
applied and tested
UMMARY OF
ELECTED
IDEO
DENTIFICATION
ETHODS AND
XPERIMENTS
ART
Method Performance Speed
MPEG-7
Video Sig.
95.49% correct detection and localization at 5ppm false alarm
(95.49% recall at 99.9995% precision)
Matching/localization: 1,500 matches/sec (pairwise matching)
Coarse signature matching: ~0.0001ms (pairwise matching)
Liu et al.
[11]
NOFA: minNDCR ~0.37 (~63% recall at 100% precision)
Balanced: minNDCR ~0.38
Mean processing time per query between 100 sec and 200 sec
Hampapur
et al. [2]
Ordinal signature: Combined recall/precision between ~94%
and 100%
Not reported
Law-To
et al. [3]
(a) Temporal ordinal signature: 100% recall at 100% precision
(b) Keypoint signature: ~82% recall at ~95% precision
Temporal ordinal: ~40 min for ~15-min query in ~3 hour DB
Keypoint: ~27 sec for ~15-min query in ~3 hour DB (DB index)
Paisitktiangk
rai et al. [6]
MAP 0.896 (DB indexing), MAP 0.911 (pairwise matching)
~20% recall at ~96% precision (with DB indexing)
0.157 sec (with DB indexing), 3.72 sec (pairwise matching)
Shen et al.
[4]
BCS: ~60% recall at ~90% precision
~20% recall at ~95% precision
50ms (with DB indexing)
Shang et al.
[5]
(a) MAP 0.953, ~20% recall at ~98% precision
(b) MAP 0.885, ~20% recall at ~98% precision
(a) 3.7ms (with DB indexing)
(b) 17ms (with DB indexing)
Sarkar et al.
[18]
MAP 0.93, ~10% recall at >99% precision
30ms (with DB indexing)
IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS FOR VIDEO TECHNOLOGY TCSVT 5350
4
producing a MAP of 0.885, with an average query speed of
17ms. However, as with [4], the ROC curves in [5] indicate
that the recall rate would drop to ~20% for a precision of
~98%. This system also does not address the issue of detecting
a video segment inside a much longer video segment.
In [6], another near-duplicate video clip detection system is
presented, which can also address the problem of detecting a
video sequence embedded in a longer video sequence with a
coarse granularity. The system samples video at a rate of 1
frame/sec and extracts signatures from 10-frame sequences
with a 50% overlap. Each clip signature is the average of its
frame descriptors, each of which is an ordinal spatial intensity
distribution (OSID). The 10-second clip signatures strike a
balance between frame-level signature approaches and video-
level signature approaches. Multi-probe locality sensitive
hashing (MPLSH) is used to index the clips. The system was
tested on the database of [5] and achieved a MAP of 0.936
and 0.947, with and without MPLSH, at the clip signature
level. At the video level, i.e. for queries comprising multiple
clip signatures the system achieved 0.896 and 0.911 MAP
with and without MPLSH respectively, with a retrieval speed
of 0.157 and 3.72 seconds per query. However, the ROC
curves in [6] also indicate that the recall rate of the system
would drop to a level of ~20% for a precision of ~96%. To
test the scalability of the system, it was tested on enlarged
versions of the original database, to ~25,000 and ~50,000
videos, with little performance degradation.
In [18] a near-duplicate video detection system is presented
which can detect temporally reduced videos, but its design
does not support detection of video content embedded in
unrelated video material. The system extracts frame-level
features based on the MPEG-7 color layout descriptor [9]. In
order to reduce the descriptor size and increase search speeds,
k-means-based clustering is performed on the frame-level
descriptors and the cluster centroids are used as the video
fingerprint. A two-stage matching approach is then used: first
k nearest neighbor matches are found using vector-quantized
fingerprint representation, distance look-up tables and dataset
pruning. Then, match candidates are confirmed or rejected
using either a distance thresholding-based approach or a
registration-based approach. Experimental evaluation was
conducted on 38,000 web videos. A subset of 1,200 videos
was used to generate 18 duplicates per video by applying
modifications including Gaussian blurring, gamma correction,
JPEG compression, frame cropping, etc. The system achieved
a MAP of ~0.93 at a query time of ~30ms. However, an
examination of the FA/FR curves in [26] shows that for higher
precision values, of ~99% or higher, the recall rate would drop
significantly and very rapidly, from ~70% to less than 10%.
As the authors note themselves, the method will not be
effective if the query contains portions of multiple videos.
Tables I and II provide a summary of the key video
identification works examined above, and also provide a
comparison with the MPEG-7 Video Signature, which will be
discussed later in this paper.
III. D
EVELOPMENT OF THE STANDARD
A. Aim and Scope
The Video Signature was designed to find identical content,
unlike previous MPEG-7 descriptors, such as Dominant Color
and Scalable Color which were designed to find semantically
similar content [9]. The scope of the Standard encompasses
those aspects which are required for interoperability. There are
four parts to the Standard, namely (i) the descriptor extraction
and decoding, along with its DDL (Descriptor Definition
Language [27]) and binary representation syntax [28], (ii) a
reference software implementation and source code for the
Video Signature Tools [29], (iii) the conditions and dataset for
ensuring conformance to the Standard [30], and (iv) an
exemplary pair-wise matching and localization scheme, as
used during the MPEG-7 evaluation process [31]. Note that
[31] is merely informative, as the Standard does not specify a
normative matching, retrieval or database indexing scheme.
Those options are left open to the application developers.
B. Requirements
The first step before the technical development of the Video
Signature Tools was the definition of the requirements, which
were subsequently used for the evaluation of the proposed
technologies. A set of 11 requirements were identified, namely
i) Uniqueness: A signature will describe a video uniquely.
ii) Robustness to editing operations: The robustness of the
Video Signature was tested against a large number of common
editing operations. These are examined in more detail later.
iii) Independence: The Video Signature shall achieve a
false alarm rate of
5 ppm
(parts per million), i.e. no more
than five false matches per million comparisons.
iv) Fast matching: This refers to the pair-wise matching
and localization speed, since the Standard does not specify any
mandatory retrieval or database indexing scheme. The
requirement was that it will be possible to match and localize
at least 1,000 clip pairs per second on a desktop PC under the
partial content matching query scenarios described later.
v) Fast extraction: This was used as a secondary criterion
in the MPEG-7 selection process, since large-scale signature
extraction is typically an offline process. In practice, the
extraction speed of the Video Signature from uncompressed
video is ~900 frames/sec on a standard desktop PC.
vi) Compactness: The compactness requirement was that
the Video Signature shall not exceed 30,720 bits/sec of
content at 30 frames/sec, i.e. 1,024 bits per frame on average.
vii) Non-alteration of the content: This ensured that the
Video Signature shall not require any content modifications,
making it suitable for immediate use with all existing content.
viii) Self-containment of the signatures: This requirement
ensured that no access to the video content from which a
signature is extracted shall be necessary for matching, making
it possible for applications to operate at the signature level.
ix) Coding independence: This requirement ensured that
the extraction of the signatures shall be independent of the
encoding of the video content.
IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS FOR VIDEO TECHNOLOGY TCSVT 5350
5
x) Partial matching: This requirement ensured that it will
be possible to detect a duplicated video segment embedded
within a longer segment of different content.
xi) Accurate temporal localization of duplicated and
embedded content: This requirement specified that the start
point/duration of duplicated and embedded content shall be
identified within one second/two seconds of the actual start
point/duration for a match to be deemed successful.
These requirements were designed to ensure that the
Standard will be applicable to many different applications.
Where hard limits are imposed, e.g. a false alarm rate of
5 ppm
or a maximum descriptor size, those were agreed
after deliberations within the MPEG-7 group on the
appropriate limits that would maximize the benefit of the
Standard in different applications and products.
IV. A
PPLICATIONS
The Video Signature has a wide range of potential
applications. Below we list a few of the main applications,
although this is not meant to be an exhaustive list.
i) Rights management and monetization: For content
owners, the aim is the detection of possible copyright
infringement or content monetization online. For content
consumers, the aim is to identify the copyright owner, if any,
to avoid infringement or ensure proper attribution.
ii) Distribution management: An organization may operate
a video fingerprint database for all its sensitive content and
automatically identify and stop accidental transmission of such
content via email, unauthorized copy to external device, etc.
iii) Usage monitoring: The aim here is to track and record
statistics such as distribution and frequency of content usage.
For example, an advertising agency may wish to check
whether its material is distributed or broadcast as expected.
iv) Video content-based linking: In a similar fashion to
text, the video content in a web page can be used to infer
association with other web pages.
v) Database management and de-duplication: For high-
volume content creators and owners, such as studios or
archives, as well as for personal video libraries.
The development process of the MPEG-7 Video Signature
Tools was guided by the requirements presented in Section III,
which were designed to include a wide range of applications.
Naturally, different applications will weigh those requirements
differently. For example, certain applications, such as personal
database management, may not require a false alarm rate of
5ppm (parts per million) or less. However, the MPEG-7 Video
Signature descriptor delivers this false alarm rate while
achieving a detection rate in the order of 96%, which is higher
than the recall rates typically reported in the literature for
significantly lower precision rates. Clearly, if a higher recall
rate is required, this can be achieved by relaxing the false
alarm rate requirement. For applications such as near-
duplicate video clip detection in large databases, the accurate
localization requirement will not apply. For such applications,
the localization aspect of the matching function may be
disabled, which will also result in significantly improved
matching speed profiles. Furthermore, in all applications
which involve a known database, suitable database indexing
may be used for the improvement of retrieval speeds.
As an International Standard, the MPEG-7 Video Signature
Tools bring two main benefits to the different systems and
applications. Firstly, MPEG follows a systematic peer-
reviewed evaluation process, leading to the adoption of the
best technologies from various proposals. Secondly, the Video
Signature Tools enable interoperability, i.e. they allow
different users and systems to talk to each other in terms of
descriptors rather than in terms of videos. This is achieved in a
number of ways. The syntax of the Video Signature is both
unambiguous and flexible. Reference software is provided as
part of the Standard, to aid in the development of compliant
products and services. Developers may also generate their own
implementations and then test their conformance according to
the conformance conditions and dataset, which also form part
of the Standard. The advantages of interoperability are
different for different parties in the content creator to
consumer chain. For example, for content creators,
standardized description tools will allow them to generate
standardized content descriptions to register their content in
different databases, while users will be able to search different
databases using the same standardized descriptors.
V. V
IDEO SIGNATURE EXTRACTION
The Video Signature comprises two parts: (i) fine signatures
extracted from individual video frames and (ii) coarse
signatures extracted from sets of frame-level signatures based
TABLE
III
K
EY
N
OTATIONS
Symbol Description Symbol Description
u
Frame signature element
ThA
j
Ternarization threshold for averages category A
j
x
Ternarized frame signature element
ThD
j
Ternarization threshold for differences category D
j
x
Vector of all 380 ternarized frame signature elements
Thd
Single Jaccard distance threshold (Matching Stage 1)
w
Set of 5 fine frame-level signature words
Thd
c
Composite Jaccard distance threshold (Matching Stage 1)
c
Fine frame-level signature confidence
Thx
H
Frame signature L1 distance threshold for Hough transform
s
Fine frame-level signature
(Matching Stage 2)
b
Coarse segment-level signature bag-of-words
Thx
G
Frame signature L1 distance threshold for matching interval
b
Set of 5 coarse segment-level signature bags-of-words
growing (Matching Stage 3)
v
Video Signature for a video
ThT
Matching segment duration threshold (Matching Stage 3)
KP / PP
Key Picture / Predicted Picture
Thc
1
Frame-level confidence threshold (Matching Stage 3)
GOP
Group of Pictures
Thc
2
Matching segment confidence threshold (Matching Stage 3)
IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS FOR VIDEO TECHNOLOGY TCSVT 5350
6
on a “bag-of-words” approach. To aid in the understanding of
the Video Signature extraction, compression and matching, a
list of the main symbols used is presented in Table III.
A. Fine Signature
The fine signatures are extracted from 32×32 pixel 8-bit
luminance information, derived by the block averaging of the
luminance channel of a frame. Each fine signature comprises
(a) a set of local features, termed the frame signature, (b) a
small representative subset of the local features, organized into
different “words” of the frame signature, and (c) a global
“frame confidencemeasure. This is illustrated in Fig. 1.
The computation of the local features involves the
calculation of local average intensities and differences at
various levels of granularity in the frame. Each local feature is
termed an element or dimension of the frame signature. In
total there are 380 such elements in a frame signature, 32
averages and 348 differences. Collectively, the frame
signature elements are designed to capture the local intensity
content and intensity interrelations at different regions and
scales of the luminance channel while being very simple to
compute. A sample of these elements is shown in Fig. 2, and
all other average and difference elements are calculated in a
similar fashion. A detailed description of all 380 elements is
not provided here, but can be found in [28].
The frame signature elements are divided into different
categories according to characteristics such as the element
type (average or difference) and pattern type. Overall, there
are two categories of average elements,
1
A
and
2
A
, with 20
and 12 elements respectively (32 average elements in total),
and eight categories of difference elements,
1
D
8
D
, with
116, 25, 36, 30, 62, 9, 50 and 20 elements respectively (348
difference elements in total). This categorization is significant
for the next step, which is the ternarization of the elements.
More specifically, the frame signature comprises ternarized
elements, i.e. quantized to three levels. We found that
ternarization produced improved results compared to
binarization, especially for the difference elements, avoiding
element value oscillation in flat frame regions. For the average
element categories
1
A
and
2
A
, this ternarization proceeds as
follows. Let
i
u
,
{ }
0,..., 1
A
j
iN∈−
, be an average element of
category
j
A
,
{ }
1, 2j
, containing a total of
A
j
N
elements.
The ternarized element
i
x
is calculated as
( )
( )
2 if 128
1 if 128
0 if 128
ij
ii j
ij
u ThA
x u ThA
u ThA
−>
= −≤
<−
(1)
The threshold
j
ThA
is not fixed but adaptive, re-calculated
for each frame and for each category
j
A
as the 33.3%
percentile rank of the absolute values
128
i
u
. That is, for
each category
j
A
with
A
j
N
elements we calculate
128
i
u
{ }
0,..., 1
A
j
iN∀∈
and sort the results in ascending order. The
threshold
j
ThA
is then the n
th
element of the sorted list, where
0.3333
A
j
nN=


. For the difference element categories
1
D
8
D
the ternarization proceeds in a similar fashion. Let
i
u
,
{ }
0,..., 1
D
j
iN∈−
, be an element of category
j
D
,
{ }
1,...,8
j
, with a total of
D
j
N
elements. The ternarized
element
i
x
is calculated as
( )
( )
2 if
1 if
0 if
ij
i ij
ij
u ThD
x u ThD
u ThD
>
=
<−
(2)
The threshold
j
ThD
is again adaptive, re-calculated for
each frame and for each category
j
D
as the 33.3% percentile
rank of the absolute values
i
u
, as described above. The aim
of making the ternarization thresholds frame-adaptive and
category-adaptive is to achieve a more uniform distribution of
the frame signature elements across the three quantization
bins. Making the thresholds frame-adaptive avoids the
situation of a frame signature with diminished information
content, e.g. when a video frame is too bright/dark or of very
poor contrast, while making them category adaptive prevents
the statistics of a certain type of element influencing the
quantization of other elements. The vector
x
of all 380
ternarized elements forms the frame signature.
In the context of the Video Signature, a “word” refers to a
compact representation of the complete frame signature and is
a small ordered subset of elements of vector
x
, i.e. a simple
projection from a 380-dimansional space to a lower-
dimensional space. For two video frames, the distance
between two corresponding words is an approximation of the
distance between the complete frame signatures. The Video
Signature utilizes
Q
Ψ
-dimensional words with
5Q =
and
5Ψ=
, i.e. five different projections from the original 380-
dimensional space of the frame signature to a 5-dimensional
space. The values
5Q =
and
5Ψ=
were found through
experimentation to provide a good tradeoff between
compactness and discriminative power. Collectively, the five
words
w
provide a good and compact representation of the
complete frame signature and are used at a later stage for the
formation of the coarse signature.
The global frame confidence value is an 8-bit integer value
calculated from the 348 difference elements of categories
1
D
8
D
. More specifically, we place all 348 absolute values
of difference elements in a single list and sort in ascending
order. Then, the 174
th
element of the sorted list, denoted by
M
, is selected as the median value and the global frame
signature confidence
c
is calculated as
( )
min 8 , 255cM=


(3)
A low confidence measure implies that the frame is a flat
x
w
c
frame signature (380 ternary elements)
words (5 words x 5 ternary elements each)
frame confidence value
s_
Fig. 1. Organization of the frame signature
x, words w
and confidence value
c
into the fine signature s.
IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS FOR VIDEO TECHNOLOGY TCSVT 5350
7
image with little or no information in the difference elements.
This information is subsequently used in the matching process.
The “frame confidence”
c
, the words
w
and the frame
signature
x
make up the fine signature
s
of the Video
Signature. It is worth noting that, in the binary representation,
the ternary elements of the frame signature and the words are
not represented by two bits each but are encoded. More
specifically, each group of five consecutive elements is
encoded into an 8-bit value, resulting in a 20% size reduction.
Thus, the fine signature described here is quite compact at
only 656 bits of storage. In practice, as part of the Video
Signature, the fine signatures require less space for storage as
temporal redundancies can be exploited, as will be seen later.
B. Coarse Signature
The coarse signatures are extracted from sets of fine
signatures based on a “bag-of-words” approach [14],[32],
[33],[34].
The bag-of-words representation is extracted for temporal
segments of 90 consecutive frames. As seen earlier, each fine
signature contains five words, i.e. five subsets of the complete
frame signature. For each of these five words, the values that it
takes over the 90 frame sequence are plotted into a 243-bin
histogram (since we have five ternary values,
5
3 243=
). Thus,
five histograms
k
h
,
{ }
1,...,5
k
, are generated, one for each
of the five subsets of the complete frame signature for each
word. Then, each histogram is binarized, by setting the value
of each bin to 1 if it is greater than 0, and leaving as 0
otherwise. This gives rise to five binary occurrence histograms
k
b
,
{ }
1,...,5k
, which become the coarse segment signature
b
for the 90-frame segment. In the Video Signature coarse
signatures are generated for 90-frame segments but with a 45-
frame overlap, e.g. if a first coarse signature is extracted for
frames
m
to
89m +
, then the second coarse signature will be
extracted for frames
45m +
to
134m +
, etc. This is illustrated
in Fig. 3. Each coarse signature requires 1215 bits for storage
which, given the coarse signature overlap, results in 810
bits/sec at 30 frames/sec.
C. Video Signature Organization
At the bitstream level, the Video Signature of a piece of
video content contains all the coarse signatures first, followed
by the fine signatures, as illustrated in Fig. 4. This
arrangement facilitates efficient bitstream access for a coarse-
(a) (b) (c) (d) (e)
(f) (g) (h) (i) (j)
Fig. 2. Representative frame signature elements: (a) One A1 element,
{ }
a
, the mean luminance of the pixels in region
a
, (b) One A2 element,
{ }
a
, (c)
Four D1 elements,
{ } { }
ab
,
{ } { }
cd
,
{ }
{ }
ef
and
{ }
{ }
gh
, (d) Two D2 elements,
{ }
{ }
ab
and
{ }
{ }
cd
, (e) Two D3 elements,
{ } { }
ab
and
{ }
{ }
cd
, (f) Two D4 elements,
{ } { }
ab
and
{ } { }
cd
, (g) Two D5 elements,
{ } { }
ab
and
{ } { }
cd
, (h) Two D6 elements,
{ }
{ }
ab
and
{ }
{ }
cd
note that the regions overlap, (i) Two D7 elements,
{ } { }
ab
and
{ }
{ }
cd
, and (j) Four D8 elements,
{ } { }
ab
,
{ } { }
cd
,
{ } { }
ef
and
{ }
{ }
gh
.
b
m
+45,
m
+134
_______
b
m
-45,
m
+44
s
m
_
s
m
+1
s
m
+44
s
m
+45
s
m
+89
b
m
,
m
+89
fine signatures_
coarse_
signatures_
s
m
+89
Fig. 3. Extraction pattern of coarse signatures b from the fine signatures s.
s
0
s
1
s
2
s
M-1
coarse_
signatures_
fine signatures_
b
0,89
b
45,134
b
M
-
90
,
M
-
1
Fig. 4. Organization of coarse signatures
b and fine signatures s in the
Video Signature bitstream for a video with
M frames.
IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS FOR VIDEO TECHNOLOGY TCSVT 5350
8
to-fine matching process, which is examined in Section VII.
VI. V
IDEO SIGNATURE COMPRESSION
The frame signatures contain a considerable amount of
temporal redundancy which is eliminated through a specially
designed lossless compression scheme. The compression is
applied only to groups of 45 consecutive frame signatures,
aligned with the coarse signature temporal segments, as
illustrated in Fig. 5. Compression is not applied to the coarse
signatures or the fine signature words or frame confidences.
Every compressed temporal segment starts with a Key
Picture (KP) frame signature. A KP frame signature is coded
in a non-predictive way. The other frame signatures of the
compressed temporal segment are Predicted Picture (PP)
frame signatures, each predicted from the temporally
preceding frame signature. A Group of Pictures (GOP) is
defined as a set of frame signatures between two KPs,
including the temporally first KP and all PPs before the next
KP. A GOP therefore consists of a KP and zero or more PPs.
The method of selecting the lengths of GOPs needs to balance
maximization of the compression performance against flexible
random access to specific frame signatures. The criterion that
is used in this work consists of counting the number of
elements that remain unchanged between the current and
reference frame signatures and comparing that number to a
predefined constant. If this number is lower than the constant,
i.e. very few elements remain unchanged, this indicates a low
correlation between the current and reference signatures, and
the current signature is encoded as a KP frame signature.
Frame signatures are transformed by modulo-3 subtraction
between corresponding elements in the current and reference
frame signature,
( )( )
, ,1 ,
mod 3
im im im
xxx
=
where
i
and
m
are the element and frame index respectively. Such
transformed frame signatures make up a prediction difference
matrix, as illustrated in Fig. 6, which is scanned into a 1-
dimensional vector. This vectorization step is done by
concatenating the columns of the prediction difference matrix
which exploits the fact that within one GOP some frame
signature elements are likely to stay constant, which leads to
longer zero runs. The structure of a GOP as coded (composed
of a KP and PPs transformed into a prediction difference
matrix) is depicted in Fig. 7.
The modulo-3 difference operation between ternary
elements of the signatures results in the three ternary values,
“0”, “1” and “2”. The vectorized prediction matrix is thus
composed of these three ternary symbols, which are coded
differently. Ternary “1” and “2” are coded with one bit each,
binary “0” for ternary “1” and binary “1” for ternary “2”. After
each ternary symbol, a run-length codeword for a run of zeros
is inserted. If no zeros follow, a codeword for zero length is
used. Ternary “0” is implicitly encoded, since after each non-
zero ternary symbol a zero run must follow, unless the non-
zero ternary symbol is last in the vectorized prediction matrix.
This coding scheme for the vectorized prediction difference
matrix is depicted in Fig. 8.
For the selection of the entropy coding method two aspects
need to be considered, namely complexity and the source
statistics. Since the presented compression method is aiming
for low-complexity implementations, we chose fixed
codeword variable length entropy coding, as it allows for
parsing of a bitstream and decoding the codewords by using a
set of very simple algorithmic steps. More specifically we
used Exponential-Golomb codes, also known as Exp-Golomb,
which offer a good approximation of the ideal code for power-
law distributions of form
~p
α
ω
βω
, where
p
ω
is the
probability of unsigned integer symbol
ω
, with constants
b___
b
c
b
b
45-frame GOPs
w
x
c w
x
c w
x
c w x
c
w x
c w x
Fig. 5. Alignment of coarse signature temporal segments with compression
temporal segments.
temporal segment frame signatures
prediction difference matrix
Fig. 6. T
emporal segment transformed into a prediction difference matrix.
Ternary “0” shown as white, ternary “1” as gray, and ternary “2” as black.
GOP_
KP
PP
coding scan
ternary “1” or “2” ternary “0”
380 ternary elements_
Fig. 7. The structure of a coded GOP.
vectrorized prediciton
matrix
run-length coded
ternary “1” or “2” ternary “0”
v
1 v 1 v 1 v
v 1
1 bit lengthvariable bit length
Fig. 8. Run
-length coding of vectorized prediction matrix.v” refer
s to
variable number of bits, while “1” refers to length of one bit.
IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS FOR VIDEO TECHNOLOGY TCSVT 5350
9
0
β
>
and
0
α
>
. These codes are a good fit for power-law
distributions that have long tails, i.e. probabilities
p
ω
are
relatively large for large
ω
, which was experimentally
confirmed to be the case for our source data [35]. Exp-
Golomb codes are parameterized by a non-negative order
number
γ
, for which the codewords are of length
( )
2
1 2 log 2l
γ
ω
ωγ

=+ +−

(4)
When selecting the order number for the Exp-Golomb
coding it was experimentally derived that
2
γ
=
provides the
best fit to the probability distribution of the zero run-lengths.
The decoding consists of inverting the operations performed
at the encoder: first Exp-Golomb and run-length decoding are
performed and the resulting ternary symbols are put into a
prediction difference matrix in the vectorization scan order.
Then, frame signatures within each GOP are processed in
order of increasing frame index
m
starting from the KP, and
then all PPs are reconstructed by computing
( )( )
, ,1 ,
mod 3
im im im
xxx
=
. Finally, each temporal segment,
potentially composed of multiple GOPs, is put in its correct
temporal position in the Video Signature. It should be noted
that, in a similar fashion to video coding standards like
MPEG-2 [36] and MPEG-4 AVC [37], it is the decoding
process of the Video Signature that forms the normative part
of the Standard. Users are free to implement their encoder in
different ways, provided that the resultant stream can be
decoded by the normative decoder.
Experiments on the MPEG-7 dataset showed that this
scheme achieves a mean compression ratio of ~23% for the
frame signature blocks, which translates to a mean
compression ratio of ~27% for the complete Video Signature.
In comparative tests, this scheme outperformed three generic
compression algorithms, namely bzip2, Prediction by Partial
Match (PPM)+Arithmetic Coding, and LZMA [38]. It is also
worth noting that, by aligning the compressed temporal
segments with the coarse signature temporal segments, and by
leaving the coarse signatures uncompressed, only a small
fraction of the compressed segments require decompression
with the coarse-to-fine matching process described next.
VII. V
IDEO SIGNATURE MATCHING AND LOCALIZATION
The Standard does not specify a mandatory matching or
retrieval scheme. For informative purposes, this section
describes the pair-wise matching and localization method that
was used during the MPEG-7 evaluation process. The
matching between two Video Signatures
1
v
and
2
v
is carried
out in three stages. The first stage uses the coarse signatures to
identify candidate matching segments. The second stage uses
the fine signatures to identify candidate parameters of frame
rate ratio and temporal offset between the candidate matching
segments. The third stage performs frame-by-frame matching
to determine candidate matching intervals using the fine
signatures, assesses the quality of each match, and selects a
best matching interval between
1
v
and
2
v
.
A. Stage 1: Coarse Signature Matching
In the first matching stage, all of the coarse signature
temporal segments of Video Signature
1
v
are compared with
all of the temporal segments of Video Signature
2
v
. For two
segments
1
f
and
2
f
, their similarity is assessed by
comparing the histograms
1
k
b
and
2
k
b
{ }
1,...,5k∀∈
and
merging the results. The distance between
1
k
b
and
2
k
b
is
measured by the Jaccard distance metric given by
12
12
1
kk
k
kk
bb
d
bb
=
(5)
This measures the distance of the segments
1
f
and
2
f
as a
function of the distinct words they have in common and all the
distinct words that they contain jointly. Since each coarse
signature
b
contains
5Q =
bags of words, we have five
Jaccard distances. These distances are fused to give the
composite distance
C
d
as
1
0
Q
Ck
k
dd
=
=
(6)
A decision on the similarity of the segments is reached by
thresholding each of the Jaccard distances
k
d
and the
composite distance
C
d
. More specifically, the segments
1
f
and
2
f
are passed to the second stage of matching if more
than half of the
k
d
Jaccard distances are less than a threshold
Thd
and the composite distance
C
d
is less than another
threshold
C
Thd
, otherwise they are declared not matching and
the matching process ends.
B. Stage 2: Temporal Parameter Estimation
For the segment pairs passed to this stage, a Hough
transform is used to identify candidate parameters of temporal
offset and frame rate ratio between the segments of each
candidate pair. These are linear properties and can therefore
be estimated using two strongly corresponding frame pairs.
First, the L1 distance between the frame signature elements
of all frame pairs between the two segments are calculated.
The frame pairs whose distance is smaller than a threshold
H
Thx
are selected as strongly corresponding frame pairs.
Then, two strongly corresponding frame pairs are selected to
calculate the temporal offset and frame rate ratio, and a vote is
cast to the calculated parameters in the Hough space. The
voting is done for all possible combinations of two strongly
corresponding frame pairs. Finally, the temporal parameter
sets with a high response in the Hough space are selected as
candidate parameter sets, and are passed to the third stage of
matching. If no parameter set in the Hough space satisfies the
selection criteria, the segment pairs are declared not matching
and the matching process ends.
C. Stage3: Localization and Verification
In the third stage of matching, the matching interval, i.e. the
start and end position of the match between the two segments,
is determined by temporal interval growing based on a frame-
by-frame matching using the frame signatures.
First, the estimated temporal offset is used to determine the
IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS FOR VIDEO TECHNOLOGY TCSVT 5350
10
initial temporal matching position. Then, using the estimated
frame rate ratio, the temporal interval is extended frame-by-
frame towards both temporal directions by calculating the L1
distance between the frame signature elements of
corresponding frames. The temporal extension stops when the
frame signature L1 distance exceeds a threshold
G
Thx
. If the
length of the resultant matching interval is shorter than a given
minimum duration
ThT
, the matching interval is eliminated as
a non-match. Otherwise, the frame confidence element
associated with each frame in the matching interval is checked
to verify the match. The overall confidence of the matching
interval is calculated as the ratio of the number of frames
which have a frame confidence that is higher than a threshold
1
Thc
to the total number of frames in the interval. If the
overall confidence is below a level
2
Thc
, the matching
interval is eliminated as a false match caused by frames with
low information content.
This process is carried out for all of the candidate temporal
parameter sets, thus generating multiple candidate matching
intervals. The best interval is finally selected as a function of
the candidate intervals’ L1 distances and lengths.
The reference software contains a complete implementation
of this matching and localization method and provides default
values for all thresholds. However, different values, or
different matching approaches, may be used in different
applications to achieve the desired operational characteristics.
VIII. MPEG-7
EVALUATION METHODOLOGY AND RESULTS
A. Dataset
The dataset used in the MPEG-7 experiments comprises
footage of various types, such as film, news, documentary,
cartoons, sport, home video, etc. The independence tests were
conducted on ~70,000 30-second queries, with a total duration
of ~570 hours, and 1,900 3-minute original, non-matching
clips, unrelated to each other, with a total duration of ~95
hours. Thus, a total of over 120 million video comparisons
were made in order to derive the required false alarm
operating settings. The robustness tests were conducted on
~70,000 30-second queries, carrying a wide range of content
modifications and with a total duration of ~600 hours, and 545
3-minute reference clips, with a total duration of ~27 hours, in
order to derive the success rates of the Video Signature.
B. Query Conditions
The Video Signature Tools were evaluated under two query
conditions, i.e. (i) direct content matching and (ii) partial
content matching. With direct content matching, the whole
query clip matches with a part of the original clip. With partial
content matching, only a part of the query clip matches with a
part of the original clip. Each of the two query types has three
query scenarios in order to evaluate performance at different
temporal granularities. The query scenarios correspond to
durations D of the segment to be matched, where D = 2, 5 and
10 seconds. In the case of partial content matching, D
indicates the minimum durations of the segment to be
matched; the total duration of the query clip is 30 seconds.
The different query scenarios are illustrated in Fig. 9.
C. Evaluation Criteria
All candidate technologies proposed to MPEG were
evaluated by setting limits for three requirements, namely (i)
independence (false alarm rate of
5 ppm
), (ii) matching
speed (at least 1,000 clip pairs per second under the partial
content matching scenarios), and (iii) descriptor size (no more
than 30,720 bits/sec of content at 30 frames/sec), and then
deciding on which technology is adopted into the Standard
based on the performance for the robustness requirement,
expressed as the mean success rate. For a given modification,
the success rate,
R
, is defined as
R CT=
where
C
is the
number of correct matches found and
T
is the number of
videos that match.
D. Independence Tests
In the independence test, each clip in a database of 1,900 3-
minute clips unrelated to each other was divided into six 30-
second clips, each of which was used to produce the required
six queries, to a total of ~70,000 queries. Each query clip was
then compared against all the unrelated 3-minute clips, to
determine the operational settings for the
5 ppm
false alarm
rate. Thus, the false alarm rate was determined based on
approximately 120 million clip comparisons.
E. Robustness Tests
In the robustness test, a database of 545 3-minute clips of
various types was used, and each clip generated the six
required query clips. Each of these queries was then subjected
to one of nine modification categories, with each category
having between one and three intensity levels, giving 22
different modifications and a collection of over 70,000 query
clips. The modified query clips were then compared with the
original clips in order to produce the detection success rates at
the operational parameters determined in the independence
test. The complete list of content modifications is shown in
Fig. 9. The MPEG
-
7 query scenarios, referred to as direct or partial
matching, with copy segment lengths of two, five and 10 seconds. Segments
in red indicate unrelated non
-reference content.
IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS FOR VIDEO TECHNOLOGY TCSVT 5350
11
Table IV. Some modifications, such as text/logo overlay, are
straightforward, while others introduce a more complex
combination of changes. For example, camera capture at SD
resolution results in a resolution change, introduction of a
border, and a significant change in color, while frame rate
reduction entails temporal interpolation at 4 frames/sec. The
MPEG-7 evaluation procedure prevented any proposed
method from anticipating a modification or otherwise using
certain characteristics of any given modification, e.g. no
proposed method could rely on the fact that frame rate
reduction was happening at known frame rates, or the
locations of the text/logo overlays.
In the MPEG-7 evaluation methodology, accurate
localization was an integral part of successful matching, i.e. a
detection was deemed successful only if it adhered to strict
localization conditions. For the direct content matching, a
match was correctly detected when the detected start point was
within one second of the actual start point. For partial content
matching, correct detection required that (i) the detected start
point in the original clip be within one second from the true
start point, (ii) the detected start point in the query clip be
within one second of the actual start point, and (iii) the
detected duration of the matching part be within two seconds
of the actual duration of the matching part.
The success rates of the Video Signature are shown in detail
in Table V, and the average success rates are summarized in
Table VI. The overall success rate of the Video Signature is
95.49%. As can be seen from Table V, the detection
performance varies with different content modifications,
strength levels, query conditions, and the length D of the
segment to be matched. The results show that, as expected,
direct content matching is a simpler problem than partial
content matching. In all cases, the length of the original video
is of no importance, as the matching process is designed to
scanthe original video in order to detect any matching
segment. The results also indicate that the length D of the
segment to be matched is of high importance, with
performance increasing as D increases. This is expected, since
the descriptors of longer segments will, in general, carry more
TABLE
IV
M
ODIFICATIONS IN THE
MPEG-7
R
OBUSTNESS
T
ESTS
Modification Type
Modification Level
Heavy Medium Light
(TLO) Text/logo overlay
30% of screen area
20% of screen area
10% of screen area
(CIF) Compression at CIF resolution
64 kbps
256 kbps
512 kbps
(RR) Resolution reduction from SD
QCIF
CIF
(FR) Frame-rate reduction from 25/30fps
4 fps
5 fps
15 fps
(CAM) Camera capture at SD resolution
10% extra border
5% extra border
0% extra border
(VCR) Analog VCR recording/recapture
3 times
2 times
1 time
(MON) Color to monochrome
I = 0.299R+ 0.587G+ 0.114B
(BR) Brightness change (additive)
+36
18
+9
(IP) Interlaced/progressive conversion
PIP or IP
TABLE
V
V
IDEO
S
IGNATURE
S
UCCESS
R
ATES
(%)
IN THE
MPEG-7
R
OBUSTNESS
T
ESTS
Modifi
cation
Direct Content Matching Partial Content Matching
D = 2 sec
D = 5 sec
D = 10 sec
D = 2 sec
D = 5 sec
D = 10 sec
H
M
L
H
M
L
H
M
L
H
M
L
H
M
L
H
M
L
TLO
79.63
89.54
98.90
87.16
93.03
99.45
88.62
93.21
100.00
66.97
71.93
78.35
65.50
75.05
78.35
65.50
73.76
97.95
CIF
99.27
99.63
99.63
99.82
100.00
99.82
99.63
100.00
100.00
96.88
99.44
98.53
97.30
99.08
98.53
97.98
99.08
98.90
RR
99.27
99.82
99.63
99.82
99.82
100.00
98.90
98.90
98.90
98.90
94.50
98.53
FR
99.63
99.27
99.63
99.63
99.82
99.82
100.00
100.00
100.00
81.18
86.11
96.34
81.89
86.51
99.44
93.73
94.62
99.72
CAM
79.27
92.27
92.29
83.67
93.55
95.05
91.01
95.21
95.60
49.17
90.61
91.01
77.25
92.82
93.94
86.42
93.55
94.50
VCR
94.42
94.81
95.78
96.45
96.85
96.88
96.59
97.04
97.24
93.63
94.07
95.96
95.68
96.11
96.70
96.25
96.48
96.69
MON
99.82
100.00
99.82
99.08
99.27
99.63
BR
100.00
98.53
99.63
100.0
99.45
99.82
100.00
100.00
99.82
98.72
98.35
98.90
98.72
98.72
98.72
98.72
98.72
99.45
IP
99.82
100.00
99.63
98.90
99.27
99.27
Modification Levels: H:Heavy M:Medium L:Light
TABLE
VI
O
VERALL
S
UCCESS
R
ATES OF
V
IDEO
S
IGNATURE
Matching
Scenario
Average Success Rate (%)
Direct
2 sec
96.65
5 sec
97.76
10 sec
98.22
Partial
2 sec
91.73
5 sec
93.43
10 sec
95.12
All
95.49
TABLE
VII
M
ATCHING
S
PEED
R
ESULTS
Matching
Scenario
Matches Per Second
Partial
2 sec
980
5 sec
1618
10 sec
2023
All
1540
IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS FOR VIDEO TECHNOLOGY TCSVT 5350
12
information, making their identification and localization less
error-prone. The durations D = 2, 5 and 10 seconds used here
are much lower than the average matching clip durations
typically reported in the literature, and were deliberately
chosen to test the performance limits of the Video Signature.
F. Speed Tests
The extraction speed of the Video Signature descriptor from
uncompressed video is ~900 frames/sec on a standard PC.
The matching speed of the Video Signature was measured
by querying a set of 1,700 3-minute clips with 300 30-second
query clips from the partial content matching tests, i.e. three
sets of 100 query clips with D = 2, 5 and 10. The average
matching speed for each of the three sets, and the overall
average, is shown in Table VII. The overall matching speed
achieved was just over 1,500 matches per second on an Intel
Xeon X5460 (single core implementation), running at 3.16
GHz and with 8GB of RAM.
Table VII shows that the matching speed increases as the
length D of the matching segment increases. Since comparing
a 30-second clip to a 3 minute clip to identify and localize a
common segment of length D is more difficult and error-prone
as D becomes smaller, this results in an increased activation
rate of Stages 2 and 3 of the matching and localization process
described in Section VII. This entails more complex
processing than Stage 1, and increased overall processing
times. Stage 1 of matching, i.e. the pair-wise matching of
coarse segment-level signatures, is deterministic in terms of
speed, and its speed is in the order of 0.0001ms per coarse
signature pair. It should again be noted that all matching speed
figures refer to pair-wise matching. No database indexing
algorithm was used in our experiments to speed up the
matching process, in order to obtain an accurate picture of the
detection performance of the Video Signature descriptor itself.
G. Storage Costs
In its uncompressed form, the fine frame-level signature is
quite compact, requiring only 656 bits of storage. The coarse
segment-level signature is also very compact, requiring only
1215 bits of storage. Thus the complete uncompressed Video
Signature storage cost is 683 bits/frame, or 20,490 bits/sec at
30 frames/sec, i.e. ~9MB per hour of video at 30 frames/sec.
In its compressed form, the complete Video Signature storage
cost is, on average, 184 bits/frame, or 5,532 bits/sec at 30
frames/sec, i.e. ~2.5MB per hour of video at 30 frames/sec.
H. Syntax Flexibility
An important decision in designing the robustness
experiments, and the Video Signature itself, was that
robustness to certain modifications should be achieved
through the flexible syntax of the Video Signature rather than
by the core descriptor. Thus, to achieve good robustness to
modifications such as picture-in-picture of arbitrary size, or
black bar insertion of arbitrary size, with or without aspect
ratio changes, one efficient approach, employed here and in
other works [11],[39], is to first detect such modifications and
then describe only the spatial regions of interest. As defined in
the Standard, a Video Signature for a given video may carry a
descriptor for the full spatial extent of a video, or for one or
more spatial sub-regions, or for the full spatial extent and for
one or more sub-regions, making it possible to address
modifications such as picture-in-picture or black bar insertion.
A number of methods exist in the literature for picture-in-
picture detection and black bar detection [11],[39], which may
be used in conjunction with the Video Signature, but are
beyond the scope of the Standard.
IX. D
ISCUSSION
Clearly, a direct comparison between the methods reported
in the literature is not straightforward for a number of reasons.
While some works address the identification of entire short
clips, others investigate the identification of a video segment
embedded in a much longer unrelated video. In the latter case,
localization may be treated as integral to correct identification,
may be reported only as auxiliary information, or may not be
reported at all. While some investigations place strong
emphasis on balancing the false positives and false negatives
in a system, others aim at extremely low false alarm rates.
This, in turn, results in diverging performance assessment
measures. Furthermore, some investigations focus on the
evaluation of entire retrieval systems, including an appropriate
database indexing component to maximize query speeds with
minimum degradation in identification performance, while
others assess the performance of video fingerprints through
pair-wise matching experiments. Finally, a like-for-like
comparison between the techniques reported in the literature is
not straightforward because of differences in the dataset sizes
and compositions, the nature of the content modifications, and
the number of queries on which results are reported.
Based on the results presented in Tables I and II, but
keeping the above factors in mind, we believe that the MPEG-
7 Video Signature compares favorably to recently reported
methods. Tested on approximately one hundred and forty
thousand queries, and with a wide range of content
modifications, the MPEG-7 Video Signature achieved an
average success rate of 95.49% with a false alarm rate no
more that 5ppm, i.e. with a precision1. This performance
relates to both the identification and the accurate localization
of a video segment embedded in a longer unrelated video
segment. In terms of matching speed, the method presented in
Section VII is, understandably, significantly slower than near-
duplicate video clip retrieval techniques, which are not
designed for accurate localization of content embedded in
longer unrelated content. However, the method described in
Section VII does not form part of the Standard, and is only
used to assess the performance of the Video Signature.
Different applications may use any matching and localization
scheme, or employ any kind of suitable database indexing
scheme, while still being conformant with the Standard. In
terms of storage costs, the uncompressed/compressed Video
Signature storage cost is very economic at 683/184 bits per
IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS FOR VIDEO TECHNOLOGY TCSVT 5350
13
frame. Clearly, global descriptors extracted at the clip level, or
descriptors which utilize temporal down-sampling of the
content, may be more economic, but will also fail to produce
the detection and localization accuracy achieved by the Video
Signature.
X. S
UMMARY AND CONCLUSIONS
The MPEG-7 Video Signature achieves high levels of
robustness to common video editing operations, and is able to
accurately detect and localize a piece of video content
embedded in a longer piece of unrelated video content, with
exceptionally low false alarm rates. The reference software
developed during the standardization process is a useful tool
for researchers and implementers and contains all the
necessary routines for the Video Signature extraction,
encoding, decoding and matching. Another essential resource
for implementers is the conformance part of the Standard,
which provides sample video data, sample Video Signature
metadata extracted by the reference software, and the
conditions that must be satisfied for any implementation to be
considered conformant to the Standard.
R
EFERENCES
[1] http://www.youtube.com/t/fact_sheet
[2] A. Hampapur, K. Hyun, and R. Bolle, “Comparison of sequence
matching techniques for video copy detection”, Conference on Storage
and Retrieval for Media Databases, pp. 194201, 2002.
[3] J. Law-To, L. Chen, A. Joly, I. Laptev, O. Buisson, V. Gouet-Brunet, N.
Boujemaa, and F. Stentiford, “Video Copy Detection: A Comparative
Study”, In Proc. 6th ACM International Conference on Image and
Video Retrieval, Amsterdam, NL, July 9-11 2007, pp. 371-378.
[4] H. T. Shen, X. Zhou, Z. Huang, J. Shao, and X. Zhou, “UQLIPS: A
Real-time Near-duplicate Video Clip Detection System”, In Proc. 33
rd
International Conference on Very Large Data Bases, Vienna, AT,
September 23-28 2007, pp. 1374-1377.
[5] L. Shang, L. Yang, F. Wang, K.-P. Chan, and X.-S. Hua, “Real-time
Large Scale Near-duplicate Web Video Retrieval”, In Proc. 2010 ACM
International Conference on Multimedia, Firenze, IT, October 25-29
2010, pp. 531-540.
[6] S. Paisitktiangkrai, T. Mei, J. Zhang, X.-S. Hua, “Scalable Clip-based
Near-duplicate Video Detection with Ordinal Measure”, In. Proc. 2010
ACM International Conference on Image and Video Retrieval, Xi’an,
CN, July 5-7 2010, pp. 121-128.
[7] ISO/IEC JTC 1/SC 29/WG 11, “MPEG Press Release”, Hannover, DE,
July 2008.
[8] ISO/IEC JTC1/SC29/WG11, “Updated Call for Proposals on Video
Signature Tools”, Busan, KR, October 2008.
[9] ISO/IEC 15938-3 2002, Information Technology Multimedia content
description interface Part 3: Visual.
[10] T. Sikora, “The MPEG-7 Visual Standard for Content Description An
Overview”, IEEE Transactions on Circuits and Systems for Video
Technology, vol. 11, no. 6, June 2001.
[11] Z. Liu, T. Liu, B. Shahraray, “AT&T Research at TRECVID 2009
Content-based Copy Detection”, 2009 TRECVid Workshop,
Gaithersburg, MD, USA, November 16-17 2009.
(http://www-nlpir.nist.gov/projects/tvpubs/tv9.papers/att.pdf)
[12] A. Joly, O. Buisson, and C. Frelicot, “Content-based copy detection
using distortion-based probabilistic similarity search”, IEEE
Transactions on Multimedia, vol. 9, no. 2, pp. 293-306, February 2007.
[13] J. Law-To, O. Buisson, V. Gouet-Brunet, and N. Boujemaa, “Robust
voting algorithm based on labels of behavior for video copy detection”,
In proc. 14th ACM International Conference on Multimedia, Santa
Barbara, CA, USA, October 23-27 2006, pp. 835-844.
[14] C. Chiu, C. Yang, and C. Chen, “Efficient and Effective Video Copy
Detection Based on Spatiotemporal Analysis”, In Proc. 9th IEEE
International Symposium on Multimedia, Taichung, TW, December 10-
12 2007, pp. 202-209.
[15] L. Chen and F. Stentiford, “Video Sequence Matching based on
Temporal Ordinal Measurement”, Pattern Recognition Letters, vol. 29,
no. 13, pp. 1824-1831, October 2008.
[16] M. Yeh and K. Cheng, “A Compact, Effective Descriptor for Video
Copy Detection”, In proc. 17th ACM International Conference on
Multimedia, Beijing, CN, October 19-23 2009, pp. 633-636.
[17] M. Bertini, A. D. Bimbo, and W. Nunziati, “Video clip matching using
MPEG-7 descriptors and edit distance”, In Proc. 5
th
ACM International
Conference on Image and Video Retrieval, Tempe, AZ, USA, July 13-
15 2006, pp. 133142.
[18] A. Sarkar, V. Singh, P. Ghosh, B. S. Manjunath, and A. Singh,
"Efficient and Robust Detection of Duplicate Videos in a Large
Database", IEEE Transactions on Circuits and Systems for Video
Technology, vol. 20, no. 6, pp. 870-885, June 2010.
[19] D. G. Lowe, “Object Recognition from Local Scale-Invariant Features”,
In proc. 7
th
International Conference on Computer Vision, Kerkyra,
GR, September 20-27 1999, vol. 2, pp. 1150-1157
[20] X. Wu, A. G. Hauptmann, and C. Ngo, “Practical Elimination of Near
Duplicates from Web Video Search,” In Proc. 15th ACM International
Conference on Multimedia, Augsburg, DE, September 2007, pp. 218
227.
[21] K. Mikolajczyk, C. Schmid, “A Performance Evaluation of Local
Descriptors”, IEEE Transactions on Pattern Analysis and Machine
Intelligence, vol. 27, no. 10, pp. 1615-1630, October 2005
[22] E. Maani, S. A. Tsaftaris, and A. K. Katsaggelos, “Local Feature
Extraction for Video Copy Detection in a Database”, In Proc. 15th
IEEE International Conference on Image Processing, San Diego, CA,
USA, October 12-15 2008, pp. 1716-1719.
[23] http://trecvid.nist.gov/
[24] A. F. Smeaton, P. Over, W. Kraaij, “Evaluation campaigns and
TRECVid”, In Proc. 8
th
ACM International Workshop on Multimedia
Information Retrieval, California, USA, October 2627 2006, pp. 321-
330.
[25] W. Kraaij, G. Awad, P. Over, “TRECVID-2009 Content-based Copy
Detection Task Overview”, 2009 TRECVid Workshop, Gaithersburg,
MD, USA, November 16-17 2009.
(http://www-nlpir.nist.gov/projects/tvpubs/tv9.slides/tv9.cd.slides.pdf)
[26] A. Sarkar, V. Singh, P. Ghosh, B. S. Manjunath, and A. Singh,
“Duplicate Video Detection: Comparison of Proposed Distance
Function with Dynamic Time Warping and Dependence of Detection
Accuracy on Keyframe Selection,” Univ. California, Santa Barbara,
CA, VRL Tech. Rep., May 2008 [Online]. Available:
http://vision.ece.ucsb.edu/publications/Sarkar−tech−report−DTW−09.p
df
[27] ISO/IEC JTC 1/SC 29/WG 11, “Introducing MPEG-7 DDL an
Overview”, Nice, FR, October 2005.
[28] ISO/IEC 15938-3:2002/AMD 4:2010, Information Technology
Multimedia content description interface Part 3: Visual, Amendment
4: Video signature tools.
[29] ISO/IEC 15938-6:2003/AMD 4:2011, Information Technology
Multimedia content description interface Part 6: Reference Software,
Amendment 4: Reference software for video signature tools.
[30] ISO/IEC 15938-7:2003/AMD 6:2011, Information Technology
Multimedia content description interface Part 7: Conformance testing,
Amendment 6: Conformance testing for video signature tools.
[31] ISO/IEC 15938-8:2002/AMD 6:2011, Information Technology
Multimedia content description interface Part 8: Extraction and use of
MPEG-7 descriptors, Amendment 6: Extraction and matching of video
signature tools.
[32] D. D. Lewis, “Naive (Bayes) at Forty: The Independence Assumption in
Information Retrieval”, In proc. 10
th
European Conference on Machine
Learning, Chemnitz, DE, April 21-24 1998, pp. 4-15.
[33] P. Scovanner, S. Ali, and M. Shah, “A 3-Dimensional SIFT Descriptor
and its Application to Action Recognition”, In proc. 15
th
International
Conference on Multimedia, Augsburg, DE, September 24-29 2007, pp.
357-360
[34] J. Sivic and A. Zisserman, “Video Google: A Text Retrieval Approach
to Object Matching in Videos”, In proc.9
th
International Conference on
Computer Vision, Nice, FR, October 13-16 2003, pp.1470-1477
IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS FOR VIDEO TECHNOLOGY TCSVT 5350
14
[35] N. Sprljan, P. Brasnett, and S. Paschalakis, “Compressed Signature for
Video Identification”, In Proc. 28
th
Picture Coding Symposium,
Nagoya, JP, December 7-10 2010, pp. 454-457
[36] ISO/IEC 13818-2 2000, Information Technology Generic coding of
moving pictures and associated audio information: Video.
[37] ISO/IEC 14496-10 2005, Information Technology Coding of audio-
visual objects Part 10: Advanced Video Coding.
[38] D. Salomon, Data Compression: The Complete Reference (4
th
Edition), Springer, pp. 139,241,851, 2007.
[39] O. Orhan, J. Liu, J. Hochreiter, J. Poock, “University of Central Florida
at TRECVID 2008: Content Based Copy Detection and Surveillance
Event Detection”, 2008 TRECVid Workshop, Gaithersburg, MD, USA,
November 17-18 2008.
Stavros Paschalakis (M’01) received the B.Eng.
degree in Computer Systems Engineering (1
st
Class
Hons.) in 1996 and the Ph.D. degree in Electronic
Engineering in 2002, both from the University of
Kent, UK. He was a Research Associate at the
University of Kent between 1997 and 1999. After
receiving his Ph.D., he joined Mitsubishi Electric
R&D Centre Europe in the UK (MERCE-UK) as a
Research Engineer, and later as R&D Manager and
Multimedia Technologies Team Leader. In 2011 he
co-founded Visual Atoms Ltd, a UK-based independent R&D and technology
consulting company. Dr. Paschalakis is the author of over 40 technical papers
and patents. His research interests include image and video processing, video
coding, multimedia databases, computer vision, and hardware
implementation of vision algorithms. He is a project editor of the MPEG-7
Video Signature Tools Standard and a chair of the MPEG-7 Visual group.
Kota Iwamoto received the B.E. degree and the
M.E. degree in Electronics, Information and
Communication Engineering from Waseda
University in 2001 and 2003, respectively. He
joined NEC Corporation in 2003 and is now an
Assistant Manager of Information and Media
Processing Laboratories. His research interests
include image processing, image/video indexing and
retrieval, image classification, and their
applications. He was involved in the activities of the
ISO/IEC JTC 1/SC 29/WG 11 (known as MPEG) for the development of
MPEG-7 Video Signature Tools standard as an ISO project editor.
Paul Brasnett received the MEng. degree in
Mathematics for Intelligent Systems in 2002 and the
PhD degree in Electrical and Electronic Engineering
in 2006, both from the University of Bristol, UK. In
2005 he joined Mitsubishi Electric R&D Centre
Europe in the UK as a Research Engineer then as a
Senior Research Engineer. In 2010 he joined
Imagination Technologies as a Senior Research
Engineer in the PowerVR Research Group. From
2007-2010 he was co-chair of the MPEG-7 Visual
group. Dr. Brasnett is an author of over 20 peer reviewed technical papers
and patents. His research interests include image and video processing, image
and video coding, computer vision, machine learning and algorithms for
general purpose GPU computing.
Nikola Sprljan received the B.Eng. degree in
Electrical Engineering at the Faculty of Electrical
Engineering and Computing at the University of
Zagreb, Croatia, in 2001 and the PhD. degree in
Electronic Engineering in 2006, at Queen Mary,
University of London, UK. He joined Mitsubishi
Electric R&D Centre Europe in the UK (MERCE-
UK) in 2006 as a Research Engineer. In 2011 he
joined Visual Atoms Ltd as a Senior Research
Engineer and then he joined Maxeler as
Acceleration Architect. His research interests include image and video
representation, media compression, streaming and robust transmission, as
well as multimedia databases and retrieval.
Ryoma Oami received the B.S. degree and the M.S.
degree from the University of Tokyo in 1993 and
1995, respectively. He joined NEC Corporation in
1995 and is now a Principal Researcher of
Information and Media Processing Laboratories. His
research interests in image and video processing and
analysis for understanding visual contents. He has
joined the standardization activities in ISO/IEC JTC
1/SC 29/WG 11 (MPEG) since 2006, and
contributed to developing several MPEG-7 and
MPEG-A standards as an ISO project editor. He is a member of the Institute
of Electrical and Electronics Engineer (IEEE).
Toshiyuki Nomura received the B.S. degree and
the M.S. degree from Nagoya University in 1990
and 1992, respectively. He joined NEC Corporation
in 1992, where he is a Research Engineer in Media
and Information Research Laboratories. His research
interests include multimedia signal processing such
as coding, analysis and understanding. He was
involved in the activities of the Audio Subgroup,
ISO/IEC JTC1/SC29/WG11 (known as
MPEG/Audio) for international standardization of
speech and audio coding. Mr. Nomura is a member of the Institute of
Electrical and Electronics Engineer (IEEE). He received the 1998 Shinohara
Memorial Academic Encouragement Award from IEICE, the 2005 Ohm
Technology Award. He is a project editor of ISO/IEC 14496-3:2001 and
ISO/IEC 14496-5:2000.
Akio Yamada received the Ph.D. degree in
information electronics science from Nagoya
University, Japan in 1993. He joined NEC
Corporation in 1993 and is now a Senior Research
Manager of Information and Media Processing
Laboratories, NEC Corp. His research interests
include visual communication systems, multimedia
signal processing such as coding, analysis and
understanding, and various media service
development. Since 1999, he has been a participant
of the MPEG and JPEG Group, where he has been active in the development
of MPEG-7, MPEG-A and JPSearch Standards as an ISO project editor. Dr.
Yamada received the ITE Niwa-Takayanagi Best Paper Award and IEICE
Yong Researcher Award in 1994 and 1996, respectively.
Miroslaw Bober (M’97) is a professor of video
processing at
the University of Surrey, UK. In 2011
he co-founded Visual Atoms Ltd, a UK-based
independent R&D and technology consulting
company. Between 1997 and 2011 he was with
Mitsubishi Electric R&D Centre Europe in the UK
(MERCE-UK) as Chief Scientist and later as
General Manager. Before joining MERCE-UK, he
was with the University of Surrey in the UK, as a
research scientist and then as lecturer and leader of
the Image Communication and Multimedia Systems Group. Dr. Bober
received an M.Sc. degree (with distinction) in Electrical Engineering from
the AGH University of Science and Technology, Poland, in 1990.
Subsequently he received an M.Sc. degree (with distinction) in Signal
Processing and Artificial Intelligence in 1991 and a Ph.D. in 1995, both from
the University of Surrey, UK. His research interests include various aspects
of image processing, computer vision and pattern recognition. He has been
involved in the development of MPEG-7 for over 10 years, chairing the
MPEG-7 visual group. Dr. Bober is an inventor of over 70 patents, and
author of over 80 refereed publications, including three books and book
chapters.
... If we select V D , there is no possibility of 0 ) , ( < j i s in Equation (1). This property would appear to be desirable because 0 is used to fill the first row and column in the local alignment of Equation (15). However, multiple levels of brightness typically appear in videos. ...
... There are many possible visual descriptors t x that can be extracted from the frame t v . Examples include MPEG-7 CSD (color structure descriptor) [13], SIFT (scale-invariant feature transform) [14], MPEG-7 video signature tools [15], and MPEG-CDVS (compact descriptor for visual search) [16] [17] that is undergoing a standardization process. They can be combined using a weighting that maintains Equation (16), such that feature vector t x resides in the simplex. ...
... However, we will check to see if this paper's method can detect the plagiarism that has more venomous intentions. That is, we inserted scenes as follows after [9] with stricter constraints. For such a set of pseudo-illegal videos, we applied following methods in the order. ...
... Step 4: Computation of precision-recall curves by the following equations: racall = # of correctly detected videos # of videos in the same class (8) precision = # of correctly detected videos # of top ranked videos to be checked (9) Note that for (9), we adopt an 11-point interpolated precision. ...
Conference Paper
Full-text available
Learning algorithms that harmonize standardized video similarity tools and an integrated system are presented. The learning algorithms extract exemplars reflecting time courses of video frames. There were five types of such clustering methods. Among them, this paper chooses a method called time-partition pairwise nearest-neighbor because of its reduced complexity. On the similarity comparison among videos whose lengths vary, the M-distance that can absorb the difference of the exemplar cardinalities is utilized both for global and local matching. Given the order-aware clustering and the M-distance comparison, system designers can build a basic similar-video retrieval system. This paper promotes further enhancement on the exemplar similarity that matches the video signature tools for the multimedia content description interface by ISO/IEC. This development showed the ability of the similarity ranking together with the detection of plagiarism of video scenes. Precision-recall curves showed a high performance in this experiment.
... The ISCC makes use of various granular fingerprinting algorithms (Phash [22], Chromaprint [23], Mpeg7 Signatures [24]). These fingerprints are the inputs to the hash-functions that create the short and content-aware binary codes that we call Content-Codes. ...
... The Video Signature Tools is an ISO/IEC standard for multimedia contents [8] that is still being developed by additions and enhancements to its toolbox. An important part is the Frame Signature, which is a luminance histogram rather than a color histogram. ...