Technical ReportPDF Available

COST292 experimental framework for TRECVID 2008


Abstract and Figures

In this paper, we give an overview of the four tasks submitted to TRECVID 2007 by COST292. In shot boundary (SB) detection task, four SB detectors have been developed and the results are merged using two merging algorithms. The framework developed for the high-level feature extraction task comprises four systems. The first system transforms a set of low-level descriptors into the semantic space using Latent Semantic Analysis and utilises neural networks for feature detection. The second system uses a Bayesian classifier trained with a "bag of subregions". The third system uses a multi-modal classifier based on SVMs and several descriptors. The fourth system uses two image classifiers based on ant colony optimisation and particle swarm optimisation respectively. The system submitted to the search task is an interactive retrieval application combining retrieval functionalities in various modalities with a user interface supporting automatic and interactive search over all queries submitted. Finally, the rushes task submission is based on a video summarisation and browsing system comprising two different interest curve algorithms and three features.
Content may be subject to copyright.
The COST292 experimental framework for TRECVID 2007
Q. Zhang1
, M. Corvaglia2
, S. Aksoy3
, U. Naci4
N. Adami2
, N. Aginako12
, A. Alatan7
, L. A. Alexandre10
, P. Almeida10
, Y. Avrithis8
J. Benois-Pineau6
, K. Chandramouli1
, U. Damnjanovic1
, E. Esen7
, J. Goya12
M. Grzegorzek1
, A. Hanjalic4
, E. Izquierdo1
, R. Jarina11
, P. Kapsalas8
I. Kompatsiaris5
, M. Kuba11
, R. Leonardi2
, L. Makris5
, B. Mansencal6
, V. Mezaris5
A. Moumtzidou5
, P. Mylonas8
, S. Nikolopoulos5
, T. Piatrik1
, A. M. G. Pinheiro10
B. Reljin9
, E. Spyrou8
, G. Tolias8
, S. Vrochidis5
, G. Yakın3
, G. Zajic9
February 26, 2008
In this paper, we give an overview of the four tasks submitted to TRECVID 2007 by COST292. In shot
boundary (SB) detection task, four SB detectors have been developed and the results are merged using
two merging algorithms. The framework developed for the high-level feature extraction task comprises
four systems. The first system transforms a set of low-level descriptors into the semantic space using
Latent Semantic Analysis and utilises neural networks for feature detection. The second system uses a
Bayesian classifier trained with a “bag of subregions”. The third system uses a multi-modal classifier
based on SVMs and several descriptors. The fourth system uses two image classifiers based on ant colony
optimisation and particle swarm optimisation respectively. The system submitted to the search task is
an interactive retrieval application combining retrieval functionalities in various modalities with a user
interface supporting automatic and interactive search over all queries submitted. Finally, the rushes task
submission is based on a video summarisation and browsing system comprising two different interest curve
algorithms and three features.
1Q. Zhang, K. Chandramouli, U. Damnjanovic, T. Piatrik and E. Izquierdo are with Dept. of
Electronic Engineering, Queen Mary, University of London, Mile End Road, London E1 4NS, UK,
2M. Corvaglia, N. Adami and R. Leonardi are with University of Brescia, Via Branze 38 25123 Brescia, ITALY,
3G. Yakın and S. Aksoy are with RETINA Vision and Learning Group, Bilkent University, Bilkent, 06800, Ankara, Turkey,
4U. Naci, A. Hanjalic are with Delft University of Technology, Mekelweg 4, 2628CD, Delft, The Netherlands,
5S. Vrochidis, A. Moumtzidou, S. Nikolopoulos, V. Mezaris, L. Makris and I. Kompatsiaris are with Informatics and
Telematics Institute/Centre for Research and Technology Hellas, 1st Km. Thermi-Panorama Road, P.O. Box 361, 57001
Thermi-Thessaloniki, Greece, {stefanos, moumtzid, nikolopo, bmezaris, lmak, ikom}
6B. Mansencal and J. Benois-Pineau are with LABRI, University Bordeaux, 351, cours de la Liberation 33405, Talence,
France, jenny.benois,
7E. Esen, A. Alatan are with Middle East Technical University, 06531, Ankara, Turkey,,
8E. Spyrou, P. Kapsalas, G. Tolias, P. Mylonas and Y. Avrithis are with Image Video and Multimedia Laboratory, National
Technical University of Athens, Athens, Greece, {espyrou,pkaps,gtolias,fmylonas,iavr}
9B. Reljin and G. Zajic are with Faculty of Electrical Engineering, University of Belgrade, Bulevar Kralja Aleksandra 73,
11000 Belgrade, Serbia,
10A. M. G. Pinheiro, L. A. Alexandre and P. Almeida are with the Universidade da Beira Interior, Covilha, Portugal,, {lfbaa,palmeida}
11R. Jarina and M. Kuba are with Department of Telecommunications, University of Zilina. Univerzitna 1, 010 26 Zilina,
Slovakia, {jarina,kuba}
12N. Aginako, J. Goya are with VICOMTech, Mikeletegi Pasealekua, 57 Parque Tecnolgico 20009 Donostia / San Sebastin,
1 Introduction
This paper describes collaborative work of several European institutions in the area of video retrieval under
a research network COST292. COST is an intergovernmental network which is scientifically completely
self-sufficient with nine scientific COST Domain Committees formed by some of the most outstanding
scientists of the European scientific community. Our specific action COST292 on semantic multi-modal
analysis of digital media falls under the domain of Information and Communication Technologies.
Being one of the major evaluation activities in the area, TRECVID has always been a target initiative
for all COST292 participants [1]. Therefore, this year our group has submitted results to all four tasks.
Based on our submissions to TRECVID last year, we have tried to improve and enrich our algorithms and
systems according the previous experience [2]. The following sections bring details of applied algorithms
and their evaluation.
2 Shot Boundary Detection Task
In shot boundary (SB) detection task, four SB detection algorithms have been developed, by the University
of Brescia (U. Brescia), the Technical University of Delft (TU Delft), the Middle East Technical University
(METU) and the Queen Mary, Univeristy of London (QMUL) respectively. These algorithms are applied
on TRECVID 2007 audiovisual contents and the results are merged using two algorithms provided by the
LaBRI, University of Bordeaux 1 (LaBRI) and U. Brescia, in order to investigate how and how much the
performance can be improved.
In the following sections, first the tools proposed by each COST292 participants are presented, then the
integration methods are described, finally the results of submission are reported.
2.1 SB Detector of University of Brescia
The algorithm13 is based on the classical twin comparison method where the error signals used to detect
transitions is based on statistical modelling. Assuming that the contents of two consecutive shots can be
represented by two independent random variables, an abrupt transition is modelled as drastic variation in
the colour density function of adjacent frames while dissolves are detected evaluating the difference between
the colour density function of the actual frame and the one predicted by the dissolve model (Figure 1).
Adaptive thresholds are used to improve the performance [3].
The detection of gradual transition is limited only to dissolves and fades using the algorithm described
in [4]. In this model, two basic assumptions are considered. The first requires that series of video frames
forming a given shot can be modelled by a stationary process, at least for a duration equal to the dissolve
length. If this hypothesis is satisfied, a single frame can be used to describe the statistics of at least a clip
of video frames. The second assumption implies the independence between the RVs describing the frames
of adjacent shots. An estimate of the marginal pdf of each process is represented by the last frame of
shotout/shotin prior to the dissolve, Fin/Fout , respectively.
If these two assumption are satisfied, the colour or luminance histogram of a frame belonging to dissolve
can be obtained as convolution of the histograms Hin and Hout properly scaled to take into account the
dissolve frame weighting. This implies that the difference between H[n] and Hin Hout should ideally be
zero. On the contrary, if Fin and Fout belong to a same shot, the previous histogram difference would be
non-zero. From this consideration, it is possible to obtain a simple criterion for dissolve detection.
In hard transition detection, in order to reduce the influence of motion, the video frames are parti-
tioned into a grid of rectangles and for each area the colour histogram is extracted. The distance between
histograms extracted from consecutive frames are then calculated. For transition detection, initially a series
on Mframes is loaded in Buffer and for each of them the distance introduced above are evaluated. All
this information are then used to adaptively estimate the thresholds used in the twin comparison detec-
tion scheme. Once the detection process is started, for each frame of the video sequence, the probability
belonging to a hard and a gradual transition are evaluated and the frames are classified accordingly. The
confidence values are provided directly as detection probability.
13The proposed method has been developed in collaboration with TELECOM Italia.
2.2 SB Detector of the TU Delft
The proposed method introduces the concept of spatiotemporal block based analysis for the extraction of
low level events. The system makes use of the overlapping 3D pixel blocks in the video data as opposed to
the many other methods that use the frames or the 2D blocks in the frames as the main processing units.
The full detail of the system can be found in [5].
The method is based on the gradient of spatiotemporal pixel blocks in video data. Derivatives in the
temporal direction ~
kand the estimated motion direction ~
vare extracted from each data block (i, j, k) of
size Cx,Cyand Ctas in the following equation.
vIi,j,k(m, n, f ) = Ii,j,k(m+vx, n +vy, f + 1) Ii,j,k(m, n, f ) (1)
where Iis the pixel intensity function and ~
v= (vx, vy), iis the estimated motion direction. We also
calculate ~
kIi,j,k(m, n, f ) where ~
k= (0,0), assuming zero motion. We calculate two different measures
from this derivative information, namely the absolute cumulative luminance change:
vIi,j,k =1
vIi,j,k(m, n, f )|(2)
and the average luminance change:
vIi,j,k =1
vIi,j,k(m, n, f )) (3)
Besides calculating the values (2) and (3), we keep track of the maximum time derivative value in a
block. For each spatial location (m, n) in the block (i, j, k), we search for the frame fmax
i,j,k (m, n) , where the
maximum luminance change takes place:
i,j,k (m, n) = argmax(|∇~
vIi,j,k(m, n, f )|) (4)
After the frames (4) are determined for each pair (m, n), we average the maximum time derivative
values found at these frames for all pairs (m, n), that is
vIi,j,k =1
vIi,j,k(m, n, f max
i,j,k (m, n))|(5)
For the detection of gradual changes two features are calculated using (2), (3) and(5):
F1(i, j, k) = max(|∇d
vIi,j,k|) (6)
F2(i, j, k) = 1 min(|∇max
vIi,j,k|) (7)
The value of F1(i, j, k) equals to 1 if the function Ii,j,k(m, n, f ) is monotonous and gets closer to zero
as the fluctuations in the function values increase. The higher the value of F2(i, j, k) (i.e. close to 1), the
more gradual (smooth) are the variations in the function Ii,j,k(m, n, f ) over time. The confidence value
for the existence of a gradual transition at any temporal interval k=Kis calculated by averaging the
F1(i, j, K)·F1(i, j, K ) over all spatial indices (i, j) at the corresponding interval K.
Detection of cuts and wipes are based on the values calculated in (4). To do this, all fmax
i,j,k (m, n) values
are fit to a plane equation and the error is calculated. Lower error values suggests an abrupt change in
the corresponding block. If the plane approximation error values are low in all blocks and the same time
index, we detect a ”cut”. On the other hand if the time indices for the planes are distributed in a short
time interval, this suggests a “wipe”.
The matrix in Figure 2 depicts the confidence values for an eight-minute sports video that contains
two cuts and two wipes. Each column depicts the values of confidences collected row by row from all
blocks sharing the same time index k. The brightness level of matrix elements directly reveals the values
of confidence. We observe that in case of a cut, high values of this feature are time-aligned, that is, they
form a plane vertical to the time axis. On the other hand, a wipe is characterized by high feature values,
which are not time-aligned, but distributed over a limited time interval.
2.3 SB Detector of the METU
Shots are classified into three groups as hard cuts, short gradual cuts consisting of three frames and
long gradual cuts. Each group is handled differently. The operations are performed on the DC image
of the luminance channel of the video frames. The DC image is preferred due to its robustness against
small changes that do not correspond to shot boundaries as well as its contribution to computation time
(depending on the video content 3-5% of real-time on an Intel Core2 Duo 2.0 GHz system).
For hard cuts Edge Difference (ED) and Pixel Difference (P D) values are utilised [6]. ED is computed
for consecutive frames by counting the differences between the edge images, which are obtained by Sobel
operator, and normalising the total difference to the number of pixels. P D is the normalised total absolute
pixel difference of consecutive frames. ED and P D values of the whole video are fed to 2-Means clustering.
Initial cluster means are experimentally chosen as (ED = 0, P D = 0) for non-hard cut and (ED =
0.18, P D = 0.1) for hard cut.
ED and PD are also used for short gradual cuts with the exception that for each frame they are also
computed using the second preceding frame in addition to the previous one, which are denoted by index
values one and two. At each frame short gradual cuts are detected by thresholding. For a short gradual
cut, P Di> τ1, ΣP Di> τ2, and EDi> τ3, where τ1,τ2, and τ3are experimentally chosen as 10, 5, and
0.2, respectively.
Long gradual cuts are found using the Edge Energy(EE), which is computed using the edge image of
the frame. Similar to [6] U-curves are searched based on the assumption that in case of a gradual cut EE
starts to decrease and attains its local minimum at the center of the gradual transition. The U-curves are
located by calculating the Least Squares estimates of the slopes of left(mL) and right(mR) lines at each
center frame using previous and next 6 frames. If mL<0, mR>0, and mRmL>1, the center of the
candidate gradual cut is located. The start and end frames of the transition are determined by searching
the frames where slopes diminishes. If the candidate is already found as a hard cut or short gradual cut,
it is discarded. Then false-positives are first eliminated by analyzing normalized Histogram (with 16 bins)
Difference(HDSE ) and Edge Difference(EDSE ) between start and end frames. For a correct gradual cut,
HDSE > τ4and EDSE > τ5, where τ4and τ5are experimentally chosen as 0.2 and 0.1, respectively.
Secondly, if motion vectors are available in the compressed bitstream, motion analysis at the center frame
is performed for further elimination. Let M1denote the total number of non-zero motion vectors, MX
denote the sum of x components of the motion vectors, and MYdenote the sum of y components of the
motion vectors. The motion parameters are computed as the average of the three consecutive frames around
the center. If M1< τ6and (|MX|+|MY|/2) < τ7, where τ6and τ7are experimentally chosen as 150 and
350, respectively, for QCIF dimension, then the candidate is settled as a long gradual cut. However, during
TRECVID 2007 tests, the motion vectors of the bitstream are discarded and not utilized during the final
2.4 SB Detector of the QMUL
Conventional shot detection methods analyse consecutive frame similarities. Therefore, most of the general
information about the shot is lost. Our work is motivated by the assumption that a shot boundary is a
global feature of the shot rather than local. General knowledge about the shot is cumulated during the
time from the information included in every frame. Information about the boundary is extracted indirectly
from the information about the interaction between two consecutive shots. Spectral clustering aggregates
contribution of each frame in the form of the objective function. By optimising the objective function when
clustering frames, individual shots are separated, and cluster bounds are used for detecting shot boundaries.
By keeping the information in the objective function, every frame of the shot has its contribution to the
overall information about the boundary. As frames that belong to the same shot are temporally aligned,
cluster borders will be points on the time axis. In our algorithm, spectral clustering algorithm Normalised
Cut [7] is used for clustering. Clustering is performed inside the sliding window, until the whole video is
analysed. Firstly, a similarity matrix is created by applying a specific similarity measure over the set of
features. Three eigenvectors corresponding to the second, third and fourth smallest egienvalues are used
for describing the structure of video inside the sliding window. Every eigenvector indicates possible choice
of the shot boundary. Specially created heuristics is then applied to analyse results of the clustering and
give final results.
in out
shot prev shot next
nout n
- n in
nout- n in
Figure 1: A possible frames weighting for dis-
solves (U. Brescia). Figure 2: An illustration of confidence values
for block based abrupt changes (TU Delft).
Figure 3: Merging framework for SBD result integration.
2.5 Merging
The merging method proposed in [2] is based on the knowledge of DB detectors in terms of performance.
Such method also assumes that the confidence values are reliable.
This year, this algorithm has been improved with the purpose of proposing a general method which
performs a kind of blind integration. In other words, we supposed to do not know the DB detectors but to
only know the SB detector results. The new merging method is based on the framework shown in Figure 3.
The results of each SB detector can be combined with the results of one of the other three SBD detectors.
The integration process requires a synchronisation process because each DB detector uses its own decoder.
If the confidence values are available and reliable, the integration is performed as follow. Let B1=b1the
set of transitions of SB detector 1 and B2=b2the set transition of SB detector 2, c1and c2the associated
confidence measures, and C1and C2two thresholds with C1< C2. If a SB b1B1does not intersect any
SB b2B2, and if c1> C1, then b1is retained as a detection result. If a SB b2B2does not intersect any
SB b1B1, and if c2> C2, then b2is retained as a detection result. In the case of b1b2,b1is retained as
a detection result. While, if the confidence value is not available or reliable, two possible integrations can
be generated: all transitions available in both B1and B2or all the transitions given by the intersection of
B1and B2.
2.6 Results
COST 292 submitted ten runs. Four runs have been submitted individually from each COST292 partici-
pants. The overall recall and precision of these runs are respectively: 0.871 and 0.762 for U. Brescia, 0.905
and 0.650 for METU, 0.727 and 0.531 for QMUL, 0.802 and 0.578 for TU Delft.
The remaining 6 submitted runs have been obtained with the merging algorithms described above. The
optimum couple of DB detectors for integration has been chosen on the training performed on TRECVID
2006 data. Since the integration was blind, two runs failed while four runs were successful. Among the four
successful runs, three have been obtained using the confidence value (SB1c,SB2c,SB3c) while one ignoring
the confidence value because it was not reliable (SB1min). The overall recall and precision obtained are
respectively: 0.795 and 0.607 for SB1c, 0.877 and 0.792 for SB2c, 0.877 and 0.792 for SB3c, 0.756 and 0.786
SB1min . We can note that the individual runs can be improved by the proposed merging method.
3 High-level feature extraction
COST292 participated to the high-level feature extraction task with four separate systems as well as
with integrated runs that combine these systems. The first system, developed by the National Technical
University of Athens (NTUA), transforms a set of low-level descriptors into the semantic space using Latent
Semantic Analysis and utilises neural networks for feature detection and is described in Section 3.1. The
second system, developed by the Bilkent University (Bilkent U.), uses a Bayesian classifier trained with a
“bag of subregions” and is described in Section 3.2. The third system, by the University of Beira Interior
(UBI), uses a multi-modal classifier based on SVMs and several descriptors and is described in Section 3.3.
The fourth system, by QMUL, uses two image classifiers based on ant colony optimisation and particle
swarm optimisation respectively, and is described in Section 3.4.
3.1 Feature extractor from NTUA
In the NTUA system, for the detection of all concepts apart from person and face, we use the following
approach [8]. We choose to extract colour and texture MPEG-7 descriptors from the keyframes and more
specifically Dominant Color,Scalable Color,Color Layout,Homogeneous Texture and Edge Histogram.
These low-level descriptions are extracted from image regions that resulted from a coarse colour segmenta-
tion. Then, a clustering algorithm is applied on a subset of the training set, in order to select a small set
of regions that will be used to represent the images. From each cluster we select the closest region to the
centroid. This region will be referred to as “region type”. Then, for each keyframe we form a model vector
description. Let: d1
i, d2
i, ..., dj
i, i = 1,2, . . . , NRand j=NC, where NCdenotes the number of region types,
NRthe number of the regions within the image and dj
iis the distance of the i-th region of the image to
the j-th region type. The model vector Dmis formed as
i}, min{d2
i}, ..., min{dNC
i}i, i = 1,2, . . . , NR.(8)
Then we follow a Latent Semantic Analysis approach. We construct the co-occurrence matrix of region
types in given keyframes of the training set. After the construction of the co-occurrence matrix, we solve
the SVD problem and transform all the model vectors to the semantic space. For each semantic concept,
a separate neural network (NN) is trained. Its input is the model vector in the semantic space and its
output represents the confidence that the concept exists within the keyframe. This model is applied to the
TRECVID video sequences to construct detectors for the following features: desert,vegetation,mountain,
road,sky ,fire-explosion,snow,office,outdoor,face and person.
The main idea of our face and person detection algorithm is based on extracting regions of interest,
grouping them according to some similarity and spatial proximity predicates and subsequently defining
whether the area obtained, represents a human body. Thus, the method initially involves detection of
salient points and extraction of a number of features representing local the colour and texture. At the next
step, the points of interest are grouped with the aid of an unsupervised clustering algorithm (DBSCAN)
that considers the density of the feature points to form clusters. In the classification stage, there is a
need for a robust feature set allowing the human form to be discriminated even in a cluttered background.
Histogram of Oriented Gradients (HoG) descriptor [9] is used to encode information associated to the
human body boundary. The method is based on evaluating well-normalised local histograms of image
gradient orientations in a dense grid. The basic idea is that local object appearance and shape can often
be characterised rather well by the distribution of local intensity gradients or edge directions, even without
precise knowledge of the corresponding gradient or edge positions. In practice, this is implemented by
dividing the image window into small spatial regions (“cells”), for each cell accumulating a local 1-D
histogram of gradient directions or edge orientations over the pixels of the cell. For better invariance to
illumination, shadowing, etc., it is also helpful to perform contrast normalisation to the local responses
before using them. Finally our human detection chain involves tiling the detection window with a dense
grid of HoG descriptors and using the feature vector in a conventional SVM based window classifier.
3.2 Feature extractor from Bilkent University
The detectors developed by Bilkent U. exploit both colour and spatial information using a bag-of-regions
representation [10].The first step is the partitioning of keyframes into regions. After experimenting with
several segmentation algorithms, we decided to use the k-means with connectivity constraint algorithm [11].
After an image is segmented into several regions, each region is modelled using the multivariate histogram
of the HSV values of its pixels with 8 bins used for the H channel and 3 bins for each of S and V channels,
resulting in a 72-dimensional feature vector. Then, a codebook of region types is constructed using the
k-means algorithm for vector quantisation. The number of codewords (k) was set to 1000 empirically. The
output of this step is a discrete type label assigned to each region.
Colour information can be very useful in discriminating objects/regions in an image if they have very
distinct colours. However, just like any other low-level features, colour cannot distinguish conceptually
different objects/regions if they fall to nearby locations in the feature space. An important element of
image understanding is the spatial information. For example, finding a region with dominant blue colour
(that may be water) and a neighbouring beige region (that may be sand) with another blue region (that
may be sky) above them can increase the possibility of being a coast image. Furthermore, two images
with similar regions but in different spatial locations can have different interpretations. Hence, spatial
information can be used to resolve ambiguities in image classification.
Different methods have been proposed to model region spatial relationships. However, it becomes a
combinatorial problem if one tries to model all possible relationships between regions in an image. Therefore,
we decided to use only the vertical relationship of “above-below” because it arguably provides a better
characterisation of the content. For example, flipping a photograph horizontally does not usually alter its
semantics but flipping it vertically or rotating it greatly perturb its perception. To determine the vertical
relative position of two regions, we use their projections on both axes. If there is an overlap between the
projections on the x-axis, their pro jections on the y-axis are compared. If they have no overlap on the
y-axis or if the overlap is less than 50 percent of the area of the smaller region, we conclude that the one
with a greater centroid ordinate is above the other one. If these overlap criteria are not met, it is concluded
that no significant vertical relative arrangement exists between these two regions. The result of this step
is a list of region pairs that satisfy the “above-below” relationship for each image.
After each region is assigned a type label and the pairwise spatial relationships are computed, each image
is represented as a “bag-of-regions”. We consider two settings for this bag-of-regions representation: 1) each
region is regarded separately and a “bag of individual regions” representation is generated, and 2) regions
that satisfy the above-below relationship are grouped together and a “bag of region pairs” representation
is constructed. Finally, these two representations are used separately to train Bayesian classifiers. Given
the positive examples for each semantic concept (high-level feature), using multinomial density models,
the probability values needed by the Bayesian decision rule are computed using the maximum likelihood
3.3 Feature extractor from UBI
UBI approach to detect high-level features is based on the detection of several descriptors that are used for
a multi-modal classification after a suitable training.
The main descriptor is the HoGs such as the ones proposed in [9]. The number of directions considered
depended on the type of object that was to be detected: either 9 or 18 directions. This description infor-
mation was complemented with colour information from the RGB colour histograms, texture information
from 9-7 bi-orthogonal filters, colour correlograms for 1 pixel distance with a colour quantisation to 16
colours, dominant colour descriptor, a combination of shape and texture information using a Scale-Space
Edge Pixel Directions Histogram.
The keyframes were subdivided into rectangular sub-images of varied size, depending on the descriptor.
These sub-images were processed individually and a classification was obtained from an SVM with RBF
kernel. The result for a shot was obtained by averaging the classification scores of each of its sub-images
in the keyframe.
3.4 Feature extractor from QMUL
The system developed by QMUL uses two image classifiers: classifier based on ant colony optimisation
(ACO) and classifier based on particle swarm optimisation (PSO).
The idea underpinning the ACO model is loosely inspired by the behavior of real ants. The real power
of ants resides in their colony brain and pheromone-driven communication within the colony. An important
and interesting behavior of ant colonies is, in particular, how ants can find the shortest paths between food
Table 1: Numerical results for high-level feature extraction.
QMUL NTUA UBI Bilkent U. Sum Product
Total true shots returned 57 56 134 646 683 629
Mean (inferred average precision) 0.002 0.001 0.004 0.010 0.014 0.011
sources and their nest. For image classification task, the ACO algorithm is implemented and it is integrated
with the semi-supervised COP-K-means approach.
In our proposal, the ACO plays its part in assigning each image to a cluster and each ant is giving its own
classification solution. Images are classified based on the probability influenced by heuristic information
and pheromone value. The main idea of finding optimal solution resides in marking classification solutions
by pheromone as follows:
τ(Xi,Cj)(t) = ρ τ(Xi,Cj)(t1) +
(Xi,Cj)(t) (9)
where ρis the pheromone trail evaporation coefficient (0 ρ1) which causes vanishing of the pheromones
over the iterations. τ(Xi,Cj)(t1) represents the pheromone value from previous iteration. τa
(Xi,Cj)(t) is a
new amount of pheromones calculated from all mants that assign image Xito the j’th cluster. Definition
of ∆τa
(Xi,Cj)(t) ensure that the pheromone increases when clusters get more apart and when each cluster
has more similar images. The ACO makes the COP-K-means algorithm less dependent on the initial
parameters and distribution of the data; hence it makes it more stable. Furthermore the ACO based
multi-modal feature mapping improves inferring semantic information from low-level feature.
PSO technique is one of the meta-heuristic algorithms inspired by Biological systems. The image clas-
sification is performed using the self organising feature map (SOFM) and optimising the weight of the
neurons by PSO [12]. The algorithm is applied to SOFM for optimising the weights of the neurons. The
objective of SOFM is to represent high-dimensional input patterns with prototype vectors that can be visu-
alised in a usually two-dimensional lattice structure [13]. Input patterns are fully connected to all neurons
via adaptable weights, and during the training process, neighbouring input patterns are projected into the
lattice, corresponding to adjacent neurons. SOFM enjoys the merit of input space density approximation
and independence of the order of input patterns. Each neuron represents an image with dimension equal
to the feature vector. Two different SOFM networks were used in detecting features. The first network
configuration is a dual layer SOFM (DL-SOFM) structure which enables training of only positive models
while the negative training models are implicitly generated by the network property. This model provides a
high degree of recall, while the second configuration is a single layer rectangular mesh (R-SOFM), enabling
explicit training of both positive and negative models. Thus enabling to achieve high precision.
3.5 Results
In addition to individual runs where the output from each system is submitted separately (QMUL: COST292R1,
NTUA: COST292R2, UBI: COST292R3, Bilkent U.: COST292R4), we combined the confidence values for
each shot for each high-level feature using the sum and product combination rules [14] and generated the
runs COST292R5 and COST292R6, respectively. The numerical results are shown in Table 1. Among
the 20 features evaluated by NIST, our detectors had relatively better success (among all submitted runs
by COST292) for the following concepts: sports, office, meeting, mountain, waterscape, animal, screen,
airplane, car, and boat/ship.
4 Interactive Search
The system submitted to the search task is an interactive retrieval application developed jointly by the
Informatics and Telematics Institute(ITI-CERTH), QMUL, University of Zilina (U. Zilina) and University
of Belgrade (U. Belgrade). It combines basic retrieval functionalities in various modalities (i.e. visual,
audio, textual) and a user friendly graphical interface, as shown in Figure 4, that supports the submission
of queries using any combination of the available retrieval tools and the accumulation of relevant retrieval
Figure 4: User interface of the interactive search platform
results over all queries submitted by a single user during a specified time interval. The following basic
retrieval modules are integrated in the developed search application:
Visual Similarity Search Module;
Audio Filtering Module;
Textual Information Processing Module;
Two different Relevance Feedback Modules.
The search system, combining the aforementioned modules, is built on web technologies, and more
specifically php, JavaScript and a mySQL database, providing a GUI for performing retrieval tasks over
the internet. Using this GUI, the user is allowed to employ any of the supported retrieval functionalities and
subsequently filter the derived results using audio and colour constraints. The retrieval results (represen-
tative keyframes of the corresponding shots) are presented ordered by rank in descending order, providing
also links to the temporally neighbouring shots of each one. The identities of the desirable shots, which
are considered as relevant to the query, can be stored by the user (Figure 4). The latter is made possible
using a storage structure that mimics the functionality of the shopping cart found in electronic commerce
sites and is always visible through the GUI. In this way, the user is capable of repeating the search using
different queries each time (e.g. different combination of the retrieval functionalities, different keywords,
different images for visual similarity search, etc.), without loosing relevant shots retrieved during previous
queries submitted by the same user during the allowed time interval. This interval is set to 15 minutes
for the conducted experiments, in accordance with TRECVID guidelines. A detailed description of each
retrieval module employed by the system is presented in the following section.
4.1 Retrieval Module Description
4.1.1 Visual similarity search
In the developed application, content based similarity search is realised using MPEG-7 visual descriptors
capturing different aspects of human perception such as colour and texture. Specifically, five MPEG-7
descriptors namely Color Layout, Color Structure, Scalable Color, Edge Histogram, Homogeneous Texture
are extracted from each image of the collection are extracted [15] and stored in a relational database. By
concatenating these descriptors a feature vector is formulated to compactly represent each image in the
multidimensional space. An r-tree structure is constructed off-line by using the feature vectors of all images
and the corresponding image identifiers. R-tree(s) [16] are structures suitable for indexing multidimensional
objects and known to facilitate fast and efficient retrieval on large scale. Principal Component Analysis
(PCA) was also employed to reduce the dimensionality of the initial space. In the query phase, a feature
vector is extracted from the example image and submitted to the index structure. The set of resulting
numbers correspond to the identifiers of the images that are found to resemble the query one. Since the
order of these identifiers is not ranked according to their level of similarity with the query example, an
additional step for ranking these images using custom distance metrics between their feature vectors is
further applied to yield the final retrieval outcome.
4.1.2 Textual information processing module
The textual query module attempts to exploit the shot audio information in the best way. This audio
information is processed off-line with the application of Automatic Speech Recognition and Machine Trans-
lation to the initial video, so that specific sets of keywords can be assigned to each shot. The text algorithm
employed by the module is the BM25 algorithm, which incorporates both normalised document length (the
associated text for every image/key-frame, in our case) and term frequency. Appropriate values for the
parameters used by BM25 have been selected as reported in [17] to produce satisfactory results. Enhancing
this approach, the module is further capable of providing related keywords to the searcher by processing the
associated text of the initial results and eventually extracting the most frequent keywords. In that way the
module receives feedback from the results and suggests additional input to the user for submitting similar
queries. Although the performance of the module is satisfactory in terms of time-efficiency, the quality of
the results greatly depends on the reliability of the speech transcripts.
4.1.3 Audio filtering tool from University of Zilina
A user has an option to use additional filtering of the search results by applying audio content based filtering
on the shots retrieved. Should a certain sound occurs in the video shots, a user has an option either to take
or omit such shots from the list of shots retrieved. The following six sound classes are defined: applause,
laugh, screaming, music, loud noise, and speech.
Sound classification approach is as follows: At first, GMM for each audio class has been trained on
our own collection of sound files (about 2 hours of audio in total). As a front-end, audio signal was
parameterized by conventional MFCCs, from which 2-D cepstral matrices are created by applying additional
cosine transform along each MFCC within 1sec. block of audio frames. One dimension of 2-D cepstrum
is quefrency and the second dimension is modulation frequency, which exposes temporal changes of each
MFCC. Audio track of each video in TRECVID 2007 collection is analysed by 2-D cepstrum in 1sec.
windows with 0.5 second shift. Then log-likehoods for all 6 GMMs are computed for each audio segment.
A segment is assigned to one of the 6 audio classes by applying the following kNN rule on log-likehood
vector space created from the labelled training data. If at least half (k/2) of the neighbours belongs to the
same class, the segment is assigned to this class, otherwise the segment is not labelled. We applied such
kNN based decision rather than maximum log-likehood decision with the aim to obtain the precision as
high as possible even if the recall may decrease. Finally, a video shot is assigned as relevant for the certain
audio class if at least 2 audio segments, within the shot, are labelled with the given sound class.
4.1.4 Relevance feedback module from QMUL
Relevance feedback (RF) scheme was initially developed for information retrieval systems in which it per-
forms an online learning process aiming at improving effectiveness of search engines. It has been widely
applied in image retrieval techniques since the 1990s. RF is able to train the system to adapt its behaviour
to users’ preferences by involving human into the retrieval process. An image retrieval framework with RF
analyse relevant or irrelevant feedback from the user and uses it to predict and learn user’s preferences. At
the mean time, more relevant image can be successively retrieved.
A system contains RF process is illustrated in Figure 5. It needs to satisfy several conditions:
Images are presented to the user for his/her feedback, but same images should not be repeated in
different iterations.
The input to the module is relevant/irrelevant information provided by the user on iterative bases.
The module should automatically learn user’s preferences by adapting the system behaviour using
the knowledge feedback from the user.
Figure 5: Generalised hybrid content-based image
retrieval systems with relevance feedback.
Figure 6: Block scheme of a CBIR sys-
tem with RF from U. Belgrade.
A general image retrieval system with RF such as the one displayed in Figure 5 can use any kind
of descriptors from low-level information of available content itself to prior knowledge incorporated into
ontology or taxonomy.
When a learning approach is considered, many kind of reasoning engine can be used to determine relevant
information. There are several common classes of RF modules such as: Descriptive models (e.g. Gaussians,
GMMs), Discriminative models (e.g. SVMs, Biased Discriminative Analyses) and Neural networks (e.g.
SOMs, Perceptrons).
In our framework, one of the RF modules is implemented by QMUL based on SVM. It combines several
MPEG7 or non-MPEG7 descriptors as a cue for learning and classification. SVM is one of the developed
supervised learning algorithms. It empirically models a system that predicts accurate responses of unseen
dataset based on limited training sets [18].
In submitted runs with QMUL RF, all experiments were conducted using linear SVM for the sake of
efficiency. Given the initial search result using visual similarity search or text-based search, users were
asked to select at least one positive and one negative examples on screen as feedback. Usually two to
five iterations were performed depending on users’ preferences, within the time limitation. Four MPEG7
descriptors: Colour Layout, Colour Structure, Edge Histogram and Homogeneous Texture and one non-
MPEG7 descriptor: Grey Level Co-occurrence Matrix were used and combined to conduct visual feature
based RF [19].
4.1.5 Relevance feedback module from University of Belgrade
In the Laboratory of digital image processing, telemedicine and multimedia (IPTM), Faculty of Electrical
Engineering, U. Belgrade, content-based image retrieval (CBIR) module with RF was derived. The module
uses low-level image features, such as colour, line directions and texture, for objective description of images.
We investigated both global visual descriptors (for a whole image) and local descriptors (for regions)
[20], and different combination of visual features mainly based on the MPEG-7 descriptors, including the
reduction of feature vector components [21]. In all cases the block scheme of U. Belgrade module was the
same, as depicted in Figure 6.
First step in any CBIR system includes the determination of relevant low-level features j= 1,2, , J,
describing as best as possible the content of each image i,i= 1,2, , I. Features are expressed by corre-
sponding numerical values, and are grouped into appropriate feature vector Fi= [Fi1, Fi1, ...Fi1] of the
length J. Feature vectors were stored in appropriate feature matrix, F=Fi, of dimension I×J. Then,
the retrieving procedure is based on relatively simple proximity measure (for instance, Euclidean distance,
Mahalanobis, or similar) between feature vectors of a query image and images from database.
Components of a feature vector matrix F=Fi=F(i, j), i= 1,2, , I ,j= 1,2, , J, are column-wised
rescaled with weighted term W1j, according to
mean(Fj)) + 2), j = 1,2, ..., J. (10)
As a similarity measure we used Mahalanobis distance metric, and as a relevance feedback strategy we
used a query shifting in combination with the Probabilistic Feature Relevance Learning (PFRL) method
[22], and the non-linear modelling capability of the radial basis functions (RBF).
From feature vectors of subjectively annotatad images as relevant Rand N, a query feature vector was
updated, by using the Rocchio’s equation:
where Fqis a previous query feature vector, ˆ
Fqis updated vector, and ¯
FRand ¯
FNare the mean values of
feature vectors of Rand Nimages, respectively. Positive constants αRand αNdetermine the influence
of Rand Nimages to query vector updating. In our work we associate a one-dimensional Gaussian RBF
with each feature vector Fifor images from database
Fq) =
exp ((Fij ˆ
Fqj )2
), i = 1,2, ..., I. (12)
The magnitude of Sirepresents the similarity between the feature vector, Fi, and the modified query ˆ
after user’s feedback. Standard deviation, σj, determines the slope of Gaussian function and, in particular,
reflects to the relevance of jth individual feature.
The functions Siare then used to determine the image similarity in a new (subjective) search process:
the magnitude of functions Siare stored in descending order, a new set of best matched images is displayed,
from which user selects and labels new relevant and irrelevant ones, thus updating RBF and refining the
search. The process is repeated until the user is satisfied with retrieved results. Usually, two to three
iterations were sufficient.
4.2 Results
We have submitted four runs to the TRECVID 2007 Search task. The four runs in our submission used
four different run types respectively. The run types and the results achieved using these runs are illustrated
in Table 2 below. It seems that the RF modules were capable of retrieving more relevant shots. However,
Table 2: Evaluation of search task results.
Run type visual
visual +
text search
visual + text search
visual + text search
+ RF U. Belgrade
of 2006
Precision out of to-
tal relevant shots
0.075 0.086 0.083 0.069 0.027
Average precision 0.098 0.110 0.096 0.078 0.023
the achieved scores were not improved due to the limitation of time and the fact that the users did the
experiments were relatively inexperienced with the interactive approaches. By analysing our results, it can
be observed that our submissions in this year generally outperformed 2006.
5 Rushes Task
The rushes task submission is based on a video summarisation and browsing system comprising two different
interest curve algorithms and three features. This system is a result of joint work of TU Delft, QMUL,
LaBRI and VICOMTech.
5.1 Interesting moment detector by TU Delft
We approach the modelling of the experience of a rushes video by extending our previous work on arousal
modelling [23]. Based on a number of audio-visual and editing features, the effect of which on a human
viewer can be related to how that viewer experiences different parts of the audiovisual material, we model
the arousal time curve that represents the variations in experience from one time stamp to another. High
arousal values ideally represent the parts of the video with high excitement, as compared to more-or-less
serene parts represented by low arousal values. The obtained curve can be used to automatically extract
the parts of the unedited video that are best capable of eliciting a particular experience in the given total
duration. We expect these high arousal parts to be more significant in the video and they should be shown
to the user in the first place.
5.2 Interesting moment detector by QMUL
The frames are firstly clustered using the Normalised Cuts algorithm, Ncut. This algorithm was first
introduced by Shi and Malik in [24] as a heuristic algorithm aiming to minimise the Normalised Cut
criterion between two sets, defined as:
NCut(A, B) = C ut(A, B ) (1/V olA + 1/V olB) (13)
where cut(A, B) = X
wi, j, and wi, j are pairwise similarities between points iand j:
wi,j =ed(i,j)2
where d(i, j) is a distance over the set of low-level feature vectors. Originally this approach was created
to solve the problem of perceptual grouping in the image data. The first step in our algorithm is to create
the similarity matrix. Instead of analysing the video on a key frame level, we use a predefined ratio of
frames in order to stimulate the block structure of the similarity matrix. The main task in the video
summarisation is to properly cluster scenes, and then to analyse clusters in order to find most informative
representatives. Spectral algorithms use information contained in the eigenvectors of data affinity matrix
to detect structures. Given a set of data points, the similarity matrix is defined as matrix Wwith elements
wi,j . Let Dbe a N×Nmatrix with values di=PjIwij,i[1, N] on its diagonal. Then Laplacian
matrix of the given dataset is defined as:
After creating the similarity matrix and solving the generalised eigensystem:
Lx =λDx (16)
with λbeing an eigenvalue and xbeing corresponding eigenvector, the next step is to determine the
number of clusters in the video, k. Automatic determination of the number of clusters is not a trivial
task. Every similarity matrix have a set of appropriate number of clusters depending on the choice of
the parameter σ. For automatic detection of number of clusters for fixed σ, we use the results of matrix
perturbation theory.The matrix perturbation theory states that the number of clusters in a dataset is highly
dependent on the stability of eigenvalues/eigenvectors determined by the eigengap , defined as:
with λiand λi+1, being two consecutive eigenvalues of (16). The number of clusters kis then found by
searching for the maximal eigengap over a set of eigenvalues:
After the number of clusters kis found, N×kmatrix Xis created by stacking the top keigenvectors
in columns. Each row of Xcorresponds to a point in the dataset and is represented in a k-dimensional
Euclidian space. Finally, kclusters are obtained by applying the K-means algorithm over the rows of X.
Results of the k-means algorithm are clusters that give importance information for various applications.
Scenes that contain different events result in non continuous clusters detected by the k-means algorithm.
These non constant clusters correspond to the scenes in the video. In order to properly detect these scenes,
frames belonging to the same clusters, separated by the frames of other clusters, should be merged in one
scene together with frames that lay between them on the time axis. This is done by analysing the structure
of the clusters obtained by the k-mean algorithm. Let I(i) be the cluster indicator of the frame i, with
i[1, N ]. First frame of the cluster Iis i1, and ibis the first frame of the cluster Iwith I(ib)6=I(ib+ 1)
and all frames between i1and ibbelonging to the same cluster. Finally clusters are merged by putting all
frames between iband ieto the same cluster, where ieis defined as:
{k|I(ib) = I(ib+k)}(19)
where ttr is experimentally determined threshold. Now each cluster is supposed to be corresponding to
one scene in the video. These scenes are further used as basic units for important event detection in the
following steps.
5.3 Features by LaBRI and VICOMTech
In rushes videos, there are some parts that are meaningless in the final edited movie and thus should not
appear in the summary. We can distinguish two kinds of such parts: unwanted frames which are generally
frames with nothing visible and unscripted parts showing the movie crew setting up a scene for example.
5.3.1 Unwanted frames
Unwanted frames are composed in particular of frames with uniform colour (often all black or gray), or
with colour bars (see Figure 7). According to our observations, such frames appear randomly during the
rushes, and we may still hear sound from the scene in the background. In order to detect these unwanted
frames, we compute a colour histogram on each channel of a frame, in RGB format. We then sum the peaks
of these histograms, and classify the frame as unwanted if this sum is superior to a given threshold. We
then use a median filter (of width 5) to filter this result, as wrong frames most often last several seconds.
For performance reason, we apply this detection only at I-frame resolution and interpolate the filtered
results for P-frames.
Figure 7: Unwanted frames: a) grey/black frame; b) sharp color bars; b) diffuse color bars.
This algorithm seems to work pretty well on totally black or gray frames and on sharp colour bars
(Figure 7 b)on the devel movies. But it is not appropriate for diffuse colour bars found on some videos.
Moreover, this method may also falsely detect some scripted scenes, very dark scenes in particular. But we
believe that scenes with so few colours are most often not very understandable and thus does not need to
be in the summary.
5.3.2 Human detection
On of the extracted features, is human detection on frames. Indeed, we believe a human is one of the more
recognisable object in a video, especially for a human summariser.
We use detection of skin colour to detect human presence in frames. Our human detection algorithm
has several steps. First, we use OpenCV [25] to detect front-facing faces, only on I frames for performance
reason. We apply a simple geometric filter to remove too small or too big faces, then a we apply a temporal
median filter on bounding boxes of detections. In a second pass, we use the algorithm described in [26].
We first train a colour detector on previous detected faces to extract a colour model for faces, specific to
this movie. then we process a second time I frames of the whole movie to detect areas corresponding to
the colour model. We apply again our temporal median filter on bounding boxes of detections. finally, we
interpolate our results to P-frame temporal resolution. The computed colour model is highly dependent of
the precision of the first step. It seems that we could improve our results with a better parametrisation of
5.3.3 Camera motion
Another extracted feature is camera motion. Indeed, camera motion is often used by the director to
highlight a significant event in a movie. So we believe that a part where a camera motion occurs is rather
important. Moreover, “camera event” is reported in the summary ground truth instructions as one of the
event to note for the summariser.
We use the algorithm described in [27]. First we estimate the global camera motion, extracting only
motion vectors from P-frames of MPEG compressed stream. Then we use a likelihood significance test of
the camera parameters to classify specific camera motions (pan, zoom, tilt).
On the provided rushes videos, it seems that many camera motions occur during unscripted parts of
scenes, during the scene set up in particular. However, this information may still be discriminatory when
other features are unavailable.
5.4 Merging and Layout
The merging of the features is performed in a heuristic manner. Once the frames in the system are
clustered, the importance of each cluster and the points in each cluster where the importance is highest
are determined. The summary of the video is prepared by merging the parts from each cluster around the
maximum importance point. The length of the cluster summaries are directly proportional to the cluster
importance values. The frame importance is a weighted sum of excitement level, detected number of faces
and camera motion type whose weights are set after user tests. The cluster importance is the average of
frame importance values in corresponding cluster.
5.5 Results
Thanks to the success of the clustering algorithm, our system performed as the second best algorithms
in terms of minimising the duplications. Also it is scored above average as easy to understand. On the
other hand, our system has performed below average in terms of fraction of inclusions. Since our system
is not designed to detect and separate events but relies on some low level properties of video to detect the
important sections, it may miss the parts with some defined events but without high excitement, camera
motion or faces.
[1] Alan F. Smeaton, Paul Over, and Wessel Kraaij. Evaluation campaigns and trecvid. In MIR ’06:
Proceedings of the 8th ACM International Workshop on Multimedia Information Retrieval, pages 321–
330, New York, NY, USA, 2006. ACM Press.
[2] J. Calic and all. Cost292 experimental framework for trecvid 2006. November 2006.
[3] Y. Yusoff, W.J. Christmas, and J.V. Kittler. Video shot cut detection using adaptive thresholding. In
BMVC00, 2000.
[4] N. Adami and R. Leonardi. Identification of editing effect in image sequences by statistical modeling.
In Picture Coding Symposium, pages 0–4, Portland, Oregon, U.S.A., April 1999.
[5] S.U. Naci and A. Hanjalic. Low level analysis of video using spatiotemporal pixel blocks. In Lecture
Notes in Computer Science, volume 4105, pages 777–784. Springer Berlin / Heidelberg, 2006.
[6] C. Petersohn. Dissolve shot boundary determination. In Proc. IEE European Workshop on the Inte-
gration of Knowledge, Semantics and Digital Media Technology, pages 87–94, London, UK, 2004.
[7] J. Shi and J. Malik. Normalized cuts and image segmentation. IEEE Transactions on PAMI, 8:888–905,
[8] E. Spyrou and Y. Avrithis. A region thesaurus approach for high-level concept detection in the
natural disaster domain. In 2nd International Conference on Semantics And digital Media Technologies
(SAMT), 2007.
[9] N. Dalal and B. Triggs. Histograms of oriented gradients for human detection. In Proceedings of the
Conference on Computer Vision and Pattern Recognition, volume 2, pages 886–893, June 2005.
[10] D. Gokalp and S. Aksoy. Scene classification using bag-of-regions representations. In Proceedings
of IEEE Conference on Computer Vision and Pattern Recognition, Beyond Patches Workshop, Min-
neapolis, Minnesota, June 23, 2007.
[11] I. Kompatsiaris and M. G. Strintzis. Spatiotemporal segmentation and tracking of objects for visu-
alization of videoconference image sequences. IEEE Transactions on Circuits and Systems for Video
Technology, 10(8):1388–1402, December 2000.
[12] Krishna Chandramouli and Ebroul Izquierdo. Image classification using self organising feature map
and particle swarm optimisation. In Proceedings of 3rd International Conference on Visual Information
Engineering, pages 313–316, 2006.
[13] T. Kohonen. The self organizing map. Proceedings of IEEE, 78(4):1464–1480, September 1990.
[14] J. Kittler, M. Hatef, R. P. W. Duin, and J. Matas. On combining classifiers. IEEE Transactions on
Pattern Analysis and Machine Intelligence, 20(3):226–239, March 1998.
[15] V. Mezaris, H. Doulaverakis, S. Herrmann, B. Lehane, N. O’Connor, I. Kompatsiaris, and M. G.
Strintzis. Combining textual and visual information processing for interactive video retrieval. In in
proceedings of TRECVID 2004, Gaithersburg, MD, USA, 2004.
[16] A. Gutmann. R-trees: a dynamic index structure for spatial searching. In proceedings of ACM
International Conference on Management and Data (SIGMOD’88), Siena, Italy, 1988.
[17] S.E. Intille and K. Sparck Jones. Simple, proven approaches to text retrieval. Technical Report
UCAM-CL-TR-356, 1997.
[18] Christopher J.C. Burges. A tutorial on support vector machines for pattern recognition. Data Mining
and Knowledge Discovery, 2(2):121–167, 1998.
[19] D. Djordjevic and E. Izquierdo. Kernel in structured multi-feature spaces for image retrieval. Elec-
tronics Letters, 42(15):856–857, 2006.
[20] S. Rudinac, M. U´cumli´c, M. Rudinac, G. Zaji´c, and B. Reljin. Global image search vs. regional search
in CBIR systems. In proceedings of Conf. WIAMIS 2007, Santorini, Greece, 2007.
[21] G. Zaji´c, N. Koji´c, V. Radosavljevi´c, M. Rudinac, S. Rudinac, N. Reljin, I. Reljin, and B. Reljin.
Accelerating of image retrieval in CBIR system with relevance feedback. Journal of Advances in
Signal Processing, 2007.
[22] J. Peng, B. Bhanu, and S. Qing. Probabilistic feature relevance learning for content-based image
retrieval. Computer Vision and Image Understanding, (1/2), 1999.
[23] L.-Q. Xu A. Hanjalic. Affective video content representation and modeling. IEEE Transactions on
Multimedia, February 2005.
[24] Jianbo Shi and Jitendra Malik. Normalized cuts and image segmentation. IEEE Transactions on
Pattern Analysis and Machine Intelligence, (8), 2000.
[25] Opencv., 2007.
[26] A. Don, L. Carminati, and J. Benois-Pineau. Detection of visual dialog scenes in video content based
on structural and semantic features. In International Workshop on Content-based Multimedia Indexing
(CBMI) 2005, L´etonie (Tampere), 2005.
[27] P. Kraemer, J. Benois-Pineau, and M. Gr`acia Pla. Indexing camera motion integrating knowledge of
quality of the encoded video. In Proc. 1st International Conference on Semantic and Digital Media
Technologies (SAMT), December 2006.
... Being one of the major evaluation activities in the area, TRECVID [1] has always been a target initiative for ITI- CERTH. In the past, ITI-CERTH participated in the search task under the research network COST292 (TRECVID 2006, 2007 and 2008 [2] [3] [4]) and in the semantic indexing (SIN) task (which is the similar to the old high-level feature extraction task) under MESH integrated project [5] (TRECVID 2008 [6]), K-SPACE project [7] (TRECVID 2007 and 2008 [8] [9]). Recently, ITI-CERTH has participated as stand alone organization in the search and high level feature tasks of TRECVID 2009 [10]. ...
Conference Paper
Full-text available
This paper provides an overview of the tasks submitted to TRECVID 2010 by ITI-CERTH. ITI- CERTH participated in the Known-item search (KIS) and Instance search (INS) tasks, as well as in the Semantic Indexing (SIN) and the Event Detection in Internet Multimedia (MED) tasks. In the SIN task, techniques are developed, which combine motion information with existing well-performing descriptors such as SIFT and Bag-of-Words for shot representation. In the MED task, trained concept detectors are used to represent video sources with model vector sequences, while a dimensionality reduction method is used to derive a discriminant subspace for recognizing events. The KIS and INS search tasks are performed with by employing VERGE, which is an interactive retrieval application combining retrieval functionalities in various modalities (i.e. textual, visual and concept search). Evaluation results on the submitted runs for the aforementioned tasks provide interesting conclusions regarding the performance of the involved techniques and algorithms.
This paper describes briefly main activities of the COST Action 292 “Semantic multimodal analysis of digital media”. This Action started in 2004 as a logical continuation and extension of previous Action 211. Results from this Action are very promissing and potentially useful for industry in the field of multimedia.
Full-text available
In this paper we investigate detection of high-level concepts in multimedia content through an integrated approach of visual thesaurus analysis and visual context. In the former, detection is based on model vectors that represent image composition in terms of region types, obtained through clustering over a large data set. The latter deals with two aspects, namely high-level concepts and region types of the thesaurus, employing a model of a priori specified semantic relations among concepts and automatically extracted topological relations among region types; thus it combines both conceptual and topological context. A set of algorithms is presented, which modify either the confidence values of detected concepts, or the model vectors based on which detection is performed. Visual context exploitation is evaluated on TRECVID and Corel data sets and compared to a number of related visual thesaurus approaches.
Conference Paper
Full-text available
This paper describes an approach to optimize query by visual example results, by combining visual features and implicit user feedback in interactive video retrieval. To this end, we propose a framework, in which video processing is performed by employing well established techniques, while implicit user feedback analysis is realized with a graph based approach that processes the user actions and navigation patterns during a search session, in order to initiate semantic relations between the video segments. To combine the visual and implicit feedback information, we train a support vector machine classifier with positive and negative examples generated from the graph structured past user interaction data. Then, the classifier reranks the results of visual search that were initially based on visual features. This framework is embedded in an interactive video search engine and evaluated by conducting a user experiment in two phases: first, we record the user actions during typical retrieval sessions and then, we evaluate the reranking of the results of visual query by example. The evaluation and the results demonstrate that the proposed approach provides an improved ranking in most of the evaluated queries.
Conference Paper
In this paper we propose a probabilistic approach for the automatic organization of collected pictures aiming at more effective representation in personal photo albums. Images are analyzed and described in two representation spaces, namely, faces and background. Faces are automatically detected, rectified and represented projecting the face itself in a common low dimensional eigenspace. Backgrounds are represented with low-level visual features based on RGB histogram and Gabor filter energy. Face and background information of each image in the collection is automatically organized by mean-shift clustering technique. Given the particular domain of personal photo libraries, where most of the pictures contain faces of a relatively small number of different individuals, clusters tend to be semantically significant beyond containing visually similar data. We report experimental results based on a dataset of about 1000 images where automatic detection and rectification of faces lead to approximately 300 faces. Significance of clustering has been evaluated and results are very encouraging.
Conference Paper
This paper investigates the role of gaze movements as implicit user feedback during interactive video retrieval tasks. In this context, we use a content-based video search engine to perform an interactive video retrieval experiment, during which, we record the user gaze movements with the aid of an eye-tracking device and generate features for each video shot based on aggregated past user eye fixation and pupil dilation data. Then, we employ support vector machines, in order to train a classifier that could identify shots marked as relevant to a new query topic submitted by new users. The positive results provided by the classifier are used as recommendations for future users, who search for similar topics. The evaluation shows that important information can be extracted from aggregated gaze movements during video retrieval tasks, while the involvement of pupil dilation data improves the performance of the system and facilitates interactive video search.
Conference Paper
Full-text available
This paper describes our work on classification of out- door scenes. First, images are partitioned into regions us- ing one-class classification and patch-based clustering al- gorithms where one-class classifiers model the regions with relatively uniform color and texture properties, and cluster- ing of patches aims to detect structures in the remaining regions. Next, the resulting regions are clustered to ob- tain a codebook of region types, and two models are con- structed for scene representation: a "bag of individual re- gions" representation where each region is regarded sep- arately, and a "bag of region pairs" representation where regions with particular spatial relationships are considered together. Given theserepresentations, sceneclassification is done using Bayesian classifiers. We also propose a novel re- gion selection algorithm that identifies region types that are frequently found in a particular class of scenes but rarely exist in other classes, and also consistently occur together in the same class of scenes. Experiments on the LabelMe data set showed that the proposed models significantly out- perform a baseline global feature-based approach.
We study the question of feature sets for robust visual object recognition, adopting linear SVM based human detection as a test case. After reviewing existing edge and gradient based descriptors, we show experimentally that grids of Histograms of Oriented Gradient (HOG) descriptors significantly outperform existing feature sets for human detection. We study the influence of each stage of the computation on performance, concluding that fine-scale gradients, fine orientation binning, relatively coarse spatial binning, and high-quality local contrast normalization in overlapping descriptor blocks are all important for good results. The new approach gives near-perfect separation on the original MIT pedestrian database, so we introduce a more challenging dataset containing over 1800 annotated human images with a large range of pose variations and backgrounds.
The self-organized map, an architecture suggested for artificial neural networks, is explained by presenting simulation experiments and practical applications. The self-organizing map has the property of effectively creating spatially organized internal representations of various features of input signals and their abstractions. One result of this is that the self-organization process can discover semantic relationships in sentences. Brain maps, semantic maps, and early work on competitive learning are reviewed. The self-organizing map algorithm (an algorithm which order responses spatially) is reviewed, focusing on best matching cell selection and adaptation of the weight vectors. Suggestions for applying the self-organizing map algorithm, demonstrations of the ordering process, and an example of hierarchical clustering of data are presented. Fine tuning the map by learning vector quantization is addressed. The use of self-organized maps in practical speech recognition and a simulation experiment on semantic mapping are discussed
The performance of shot detection methods in video sequences can be im-proved by the use of a threshold that adapts itself to the sequence statistics. In this paper we present some new techniques for adapting the threshold. We then compare the new techniques with an existing one, leading to an im-proved shot detection method.
Article Most of the current image retrieval systems use “one-shot” queries to a database to retrieve similar images. Typically a K-nearest neighbor kind of algorithm is used, where weights measuring feature importance along each input dimension remain fixed (or manually tweaked by the user), in the computation of a given similarity metric. However, the similarity does not vary with equal strength or in the same proportion in all directions in the feature space emanating from the query image. The manual adjustment of these weights is time consuming and exhausting. Moreover, it requires a very sophisticated user. In this paper, we present a novel probabilistic method that enables image retrieval procedures to automatically capture feature relevance based on user's feedback and that is highly adaptive to query locations. Experimental results are presented that demonstrate the efficacy of our technique using both simulated and real-world data.