ACTIVE, an Extensible Cataloging Platform for Automatic Indexing of
Audiovisual Content
Maurizio Pintus, Maurizio Agelli, Felice Colucci, Nicola Corona, Alessandro Sassu
and Federico Santamaria
CRS4, Loc. Piscina Manna, Edificio 1, Pula, Italy
Face Detection, Face Recognition, Speaker Recognition, Caption Recognition, Digital Asset Management.
The cost of manual metadata production is high, especially for audiovisual content, where a time-consuming
inspection is usually required in order to identify the most appropriate annotations. There is a growing need
from digital content industries for solutions capable of automating such a process. In this work we present
ACTIVE, a platform for indexing and cataloging audiovisual collections through the automatic recognition of
faces and speakers. Adopted algorithms are described and our main contributions on people clustering and
caption-based people identification are presented. Results of experiments carried out on a set of TV shows and
audio files are reported and analyzed. An overview of the whole architecture is presented as well, with a focus
on chosen solutions for making the platform easily extensible (plug-ins) and for distributing CPU-intensive
calculations across a network of computers.
Digital assets have pervaded all segments of mod-
ern economies, stimulating the development of a wide
range of technological platforms for cataloging, orga-
nizing and preserving large digital collections. How-
ever, a considerable amount of effort is still taken by
metadata production. This is particularly evident with
audiovisual content, where a time-consuming visual
inspection is usually required for indexing the rele-
vant parts of the video timeline.
The objective of this work is to describe a platform
that was created with the express purpose of automat-
ing the indexing process of audiovisual material.
The platform was developed with the aim of pro-
viding an intelligent cataloging infrastructure capable
of adding value to audiovisual archives of the digital
content industry, both at production and distribution
In order to narrow such a huge scope, the index-
ing has been circumscribed to the retrieval of people,
through the automatic recognition of faces, captions
and speakers. However, the paper also describes the
adopted approach for allowing the platform to be eas-
ily extended to include new indexing algorithms and
An overview of the algorithms (and of how they
fit in the whole indexing process) is provided. The
main contributions of our work are illustrated, which
are: (1) a people clustering method based on face and
clothing information; (2) a people recognition method
based on extracting names from captions overlaid on
the video frames. Results of experiments carried out
on a set of TV shows are presented and analyzed.
1.1 Outline
Section 2 will provide an architectural overview of
the ACTIVE platform in terms of its main compo-
nents.The adopted approaches for making the plat-
form extensible through plug-ins and for distributing
processing on many workers will be also described in
this section.
Section 3 will provide an overview of the platform
Section 4 will describe the algorithms and tech-
niques used to automatically index video content by
recognizing faces, as well as the results of an experi-
mentation carried out on a set of TV programs.
Section 5 will describe the algorithms and tech-
niques used to automatically index audio content by
recognizing speakers, as well as the results of an ex-
perimentation carried out on a set of audio files.
Finally, section 6 will provide a summary of the
results of the project and will highlight the implica-
tions that these results may have in terms of various
Pintus, M., Agelli, M., Colucci, F., Corona, N., Sassu, A. and Santamaria, F.
ACTIVE, an Extensible Cataloging Platform for Automatic Indexing of Audiovisual Content.
DOI: 10.5220/0005722205740581
In Proceedings of the 11th Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2016) - Volume 4: VISAPP, pages 574-581
ISBN: 978-989-758-175-5
2016 by SCITEPRESS Science and Technology Publications, Lda. All rights reserved
Figure 1: Schema of the ACTIVE System architecture.
application scenarios.
The ACTIVE Platform is based on a modular archi-
tecture, which was designed with the express inten-
tion of enabling a wide range of applications.
As shown in figure 1, the architecture defines three
main blocks: the ACTIVE Framework, the Applica-
tion Tier and the ACTIVE Extensions.
The ACTIVE Framework includes a set of
reusable components that address the requirements of
a variety of applications. These components are:
1. the ACTIVE Core, which provides a set of basic
features for managing digital assets;
2. the Job Manager, which allows to distribute
CPU-intensive processing onto several machines;
3. a Network Storage for audiovisual resources,
based on NFS.
The ACTIVE Core implements a data model for
users, digital items (internal representation of digital
assets and their metadata) and tags (bindings between
items and entities, with the latter representing key-
words, people, or other kinds of objects which may
be defined in future releases). A tag may also be asso-
ciated to a set of temporal intervals (Dynamic Tags),
which specify the time slices of the audiovisual item
the entity actually refers to.
A plug-in system provides a simple mechanism
for extending the ACTIVE Core with server-side
scripts. The whole of plug-ins builds up a reusable
codebase (ACTIVE Extensions) that other applica-
tions can benefit from.
The ACTIVE Core exposes a REST API which
allows applications and plug-in scripts to access the
internal data models using CRUD operations.
The Job Manager allows to distribute asyn-
chronous jobs across several computers, in order to
reduce the execution time of CPU-intensive opera-
tions (e.g. transcoding, image processing).
The Application Tier contains the modules that
provide the overall experience and functionality for
any system based on the ACTIVE Framework. These
modules can include any kind of applications obeying
the requirements of the final system. In the specific
case of the platform described in the present paper,
the application tier contains three distinct web appli-
cations: the main user interface, the navigator and the
The Main User Interface provides a very basic
ACTIVE, an Extensible Cataloging Platform for Automatic Indexing of Audiovisual Content
(a) (b)
Figure 2: (a) The Navigator. (b) The Summarizer.
set of digital asset management operations, allowing
to carry out simple tasks such as uploading new items,
inspecting and editing metadata, launching plug-in
scripts and applications, performing searches and dis-
playing search results in a thumbnail grid.
The Navigator (figure 2) is a tool for visually
browsing the Dynamic Tags of an item and for do-
ing some basic editing of the tags associated to people
(e.g. to correct the names in case of misrecognition).
The Summarizer (figure 2) is a tool for evaluating
the results of a search operation based on a specific
person. It provides a visual and interactive represen-
tation of the occurrences of the query term in differ-
ent temporal sections of all video items returned as
search results. Short samples from these occurrences
are concatenated in a unique timeline, so that the user
can enjoy an overview of the relevant parts of search
A short description of the workflow is given, high-
lighting the aspects related to face / speaker extrac-
tion. In order to carry out these tasks, audiovisual re-
sources shall be first imported into the ACTIVE plat-
form. As soon as the import is completed, embedded
metadata are extracted, previews are calculated and
the newly created items are indexed (so, they may ap-
pear in search results).
Face and speaker extraction can be manually
launched at any time and their progress can be moni-
tored through the Job Monitor. As soon as the extrac-
tion has been completed, a set of tags (with associated
dynamic tags) are created and the item is indexed by
them. It is also possible to use the Navigator to in-
spect the face/speaker extraction results, both in terms
of clusters (set of time slices where the same person is
assumed to be present) and identities (labels assigned
to each cluster).
The face/speaker recognition operates on the ba-
sis of previously built face/speaker models. In case a
detected face (or speaker) does not match any model,
it is assumed to belong to a new person and is labeled
as “UNKNOWN XX”, where XX is a unique string.
Assigned labels can be manually edited, either for as-
signing an identity to people originally labeled as un-
knowns, or for correcting the labels assigned to mis-
recognized people. Although a manual editing is re-
quired in case face (or speaker) recognition fails, this
editing is automatically applied to the whole cluster,
allowing significant time and effort saving.
In figure 3 the schema of our visual indexing system
is shown. On the basis of the results by (Korshunov
and Ooi, 2011), all videos are analyzed at a frame rate
of about 5 fps.
Considered frames are grouped into shots, using
the local thresholding method presented in (Dugad
et al., 1998). Histogram calculation is carried out in
the HSV space, not considering pixels with high H
values and low S and V values. Thereafter frames are
analyzed in order to find all faces in them. Face de-
tection is based on the OpenCV (Bradski, 2000) im-
plementation of a method initially proposed by Vi-
ola and Jones (Viola and Jones, 2001) (Viola and
Jones, 2004) and improved by Rainer Lienhart (Lien-
hart et al., 2003).
Found faces are then tracked, in order to obtain
face tracks belonging to the same person, using a
method based on the OpenCV implementation of the
Continuously Adaptive Mean Shift (CAMSHIFT) al-
gorithm (Bradski, 1998), and aligned.
In the next step, a people clustering block aggre-
gates face tracks that are likely to belong to the same
person by using face and clothing information. After
this phase is completed, people recognition is carried
out by using face and caption information.
VISAPP 2016 - International Conference on Computer Vision Theory and Applications
Figure 3: Schema of our visual indexing system.
4.1 People Clustering
People clustering merges found face tracks into clus-
ters in which the same person should be visible: this
is useful both for an efficient annotation by users and
for improving people recognition. People clustering
often relies on features from faces and clothing (Berg
et al., 2004) (Sivic et al., 2006) (Everingham et al.,
2006) (Maji and Bajcsy, 2007) (Zhang et al., 2009)
(El-Khoury et al., 2010).
In our system, firstly face models and clothing
models are calculated for each face track, considering
only detected faces (i.e. not considering faces located
only by tracking). LBP histograms (Ahonen et al.,
2004), calculated on the equalized aligned faces, are
used as face features. We used the LBP
with a 4 x 8 grid (in order to have square cells with the
chosen alignment, that produces face images a size of
200 x 400 pixels.). HSV color histograms from a re-
gion just below each face detection bounding box are
used to represent clothes. Width of clothing bound-
ing box is 2 times the width of the face bounding box,
while its height equals the height of the face bound-
ing box. This region permits to include the most sig-
nificant zone for clothing comparison, e.g., the mask
presented in (Sivic et al., 2006), while including lit-
tle background and few occluding objects as hands or
Face tracks are scanned in a sequential way. In the
first scan, the first face track is labeled as “Person 1”
and the other N - 1 face tracks are compared to it. If a
face track turns out to be similar to the first face track,
it is considered belonging to the same person and it is
also labeled as “Person 1”. After the first face track
has been compared to all the other face tracks, the first
face track still not labeled is labeled as “Person 2”
and the remaining not labeled face tracks are labeled
according to it. The process is repeated until all face
tracks are labeled. Eventually, the face tracks with the
same label are merged into the same cluster. Because
faces that are visible at the same time should belong
to different people, if two face tracks overlap in time
they are considered belonging to different people and
are not compared.
When comparing two face tracks, face track 1 and
face track 2, firstly the minimum χ
distance (d
f ace
between the LBP histograms from the faces in face
track 1 and the LBP histograms from the faces in face
tracks 2 , is calculated.
Having fixed two thresholds, th
f acelow
f acehigh
, with th
f acelow
f acehigh
, the two face
tracks are considered belonging to the same person if
f ace
< th
f acelow
or if all the following conditions
are verified:
1. th
f acelow
f ace
< th
f acehigh
2. All clothing bounding boxes in the two face tracks
are entirely contained by the respective frames;
3. The minimum χ
distance between the color his-
tograms from the clothes in the two face tracks is
below a local threshold, th
, calculated from
the two face tracks.
Local threshold for clothes comparison is calculated
in the following way:
= max
i, j
where i and j are indexes of histograms belonging to
face track 1, while m and n are indexes of histograms
belonging to face track 2. So d
i, j
represents the χ
distance between a histogram i and a histogram j both
belonging to face track 1; d
represents the χ
tance between a histogram m and a histogram n both
belonging to face track 2; l
and l
are the numbers of
histograms considered in face track 1 and 2, respec-
ACTIVE, an Extensible Cataloging Platform for Automatic Indexing of Audiovisual Content
4.2 People Recognition
At this stage, an association of people clusters with
real names is carried out. People recognition is usu-
ally based on feature extraction from faces. However,
other types of features have been used, like subtitles
and transcripts (Everingham et al., 2006), film scripts
(Zhang et al., 2009), strings extracted from captions
and audio tracks (Bertini et al., 2001).
In our system, firstly an attempt to use caption
recognition for labeling clusters is made. Secondly, in
those cases where this strategy turns out unsuccessful,
face recognition is used to label people clusters.
We indicate with the term caption recognition the
analysis of text overlaid on video frames in order to
find the names of the visible people. A database (tag
dictionary) shall be prepared, containing a list of tags
identifying the names of the people which may ap-
pear in the videos. Words extracted from frames are
matched with these tags. A reduced bitrate of 1 frame
per second is used in this case and only frames with
one face in them are considered. The tool used for
OCR is tesseract (Smith, 2007); it is set to recognize
only letters, both lowercase and uppercase.
Original frame is binarized by using Otsu’s
method (Otsu, 1979), then all contours in image are
retrieved. Contours that are too small or too big are
discarded, the remaining ones are analyzed by the
OCR engine in order to recognize single characters.
Found characters are ordered in rows by checking
their bounding boxes, discarding characters that are
inside other ones. For each row, the portion of orig-
inal image that contains all characters in the row is
binarized by using Otsu’s method and found charac-
ters are put in another binary image that is analyzed
by the OCR engine in order to recognize words in it.
Each found caption block is matched against tags
from the tag dictionary by using the Levenshtein dis-
tance (Levenshtein, 1966), obtaining, for each tag, a
similarity measure between 0 and 1. If the tag that
gets the maximum similarity measure is greater than
a given threshold, it is assigned to the face. Only tags
that are assigned to at least 4 frames in the cluster are
considered (usually, a caption is visible for at least 4
seconds). If certain words are identified in a frame,
e.g. indicating that the overlaid name refers to a per-
son speaking on the phone, the frame is not consid-
Face comparison is the same used in people clus-
tering, with the difference that faces in the video are
compared with previously built face models. In both
caption recognition and face recognition, results from
single frames are aggregated with a majority rule in
order to obtain a final tag for the cluster.
4.3 Experimental Results
Experiments on our visual indexing system were car-
ried out on three full-length episodes of three different
TV shows from the Sardinian channel “Videolina”,
“Facciamo i conti”, “Monitor” and “SportClub sugli
Spalti”, with durations of about 54, 119 and 172 min-
utes. All used videos have a resolution of 720 x 576
pixels. Only people that could be relevant for a user
were considered (e.g., pedestrians were ignored), re-
spectively 6, 10 and 15 for the three videos: start
times and durations of video segments in which these
people are visible were manually annotated. Anno-
tations extracted by the system were then compared
with these ones.
Performance is measured by averaging precision,
recall and f-measure calculated on considered people
in the video. For each person p we have:
F measure
= 2 ·
· recall
+ recall
where T
is the total duration of video segments
correctly assigned to person p, T
is the total du-
ration of video segments assigned to person p, T
is the total duration of video segments in which that
person is actually visible.
Firstly, experiments on people clustering, without
the use of people recognition, were carried out. In
this case, for each cluster, the central frame of the
first face track was selected as a keyframe, highlight-
ing the face of the considered person: annotation was
carried out semi-automatically by assigning tags to
the clusters according to visual inspection of these
keyframes. Figure 4 reports experimental results for
several values of threshold for face comparison.
Increasing the thresholds, more face tracks are
merged and so the number of detected clusters de-
creases. Using only face features, average F-measure
remains almost constant for low threshold values and
exhibits a sharp decrease when the threshold is above
a critical value, more or less equal to 8. When cloth-
ing features were used, th
f acelow
was fixed to 8 and
only th
f acehigh
was changed. In the two longest
videos, the combined use of face and clothing features
outperforms the use of only face features, reducing
the decrease of F-measure.
Then, experiments on people recognition were
carried out; in this case a fully automatic annotation
was used. A threshold just below the critical value
VISAPP 2016 - International Conference on Computer Vision Theory and Applications
(a) (b) (c)
Figure 4: People clustering results. The mean values of F-measure are reported, with and without the use of clothing features,
for three full-length episoded of three TV shows,“Facciamo i conti” (a), “Monitor” (b) and “SportClub sugli Spalti” (c).
(a) (b) (c)
Figure 5: People recognition results. The mean values of precision and recall are reported, using only caption recognition,
only face recognition and using both, for three full-length episodes of three TV shows,“Facciamo i conti” (a), “Monitor” (b)
and “SportClub sugli Spalti” (c).
was chosen for people clustering. Training set for the
creation of the face models used by face recognition
was built from frames extracted from 15 videos be-
longing to the three TV shows (including those used
in the experiments). Face detection and face align-
ment was automatically performed, so these frames
were cropped, if necessary, to include only one per-
son. 12 images per person per each of 80 people
were considered; similarly to the PIE database (Sim
et al., 2002), each person was present under 3 differ-
ent poses (frontal, turned left and turned right) and
with 4 different expressions (neutral, smiling, blink-
ing and talking). The names of these 80 people consti-
tuted the tag dictionary used by caption recognition.
Figure 5 reports experimental results.
In the first two considered TV shows the use of
captions provides good results, while in “SportClub
sugli Spalti”, where there is a lot of text overlaid
on frames (sport results and standings), results using
captions are worse. Best results are obtained using
thresholds of 0.7-0.9 for caption recognition and 10
for face recognition.
The audio indexing system performs speaker recog-
nition on audiovisual items. The system, whose
schematic diagram is outlined in figure 6, is based on
the LIUM Speaker Diarization framework (Meignier
and Merlin, 2010) (Rouvier et al., 2013).
The key concepts of the system are summarized
1. the extracted audio is classified to obtain seg-
ments containing either music, silence or speech
(Ajmera et al., 2004);
2. the speech parts are segmented on the basis of
speaker changes;
3. segments are grouped into clusters, each one con-
taining speech from the same speaker;
4. the speaker recognition process associates to each
cluster an audio tag, representing the name of the
A two-step speaker diarization has been used (Barras
et al., 2006). First, in order to produce homogeneous
speech segments, an acoustic BIC (Bayesian Informa-
tion Criterion) segmentation (Chen and Gopalakrish-
nan, 1998) is carried out, using generalized likelihood
ratio as metric (Meignier and Merlin, 2010) to deter-
mine the similarity over an audio segment, followed
by a BIC hierarchical clustering, which groups similar
segments. Next, a Viterbi resegmentation is applied,
which produces a new set of segment clusters.
The next step is the the Speaker Recognition pro-
cess, which is based on the extraction of a set of fea-
ACTIVE, an Extensible Cataloging Platform for Automatic Indexing of Audiovisual Content
Figure 6: Schema of our audio indexing system.
tures (Khan et al., 2008).
The voice of an unknown speaker is analyzed and
modeled as a random Gaussian process from which
its corresponding sequence of MFCC vectors is ex-
tracted. The parameters of the Gaussians are com-
puted from MFCC vectors, using a Maximum Like-
lihood (ML) method. The model of the unknown
speaker is compared with the model of known speak-
For accomplishing this task our system uses a sta-
tistical background model from which the models of
each speaker are adapted (Reynolds et al., 2000). The
specific models of each speaker are adapted from the
UBM using the maximum a posteriori (MAP) estima-
tion (Gauvain and Lee, 1994).
5.1 Experimental Results
Experiments were carried out on a set of audio files
with the following characteristics:
high-quality audio (40 files of around 20 minutes
each, one speaker for each file, 5 total speakers);
average-quality audio (40 files of around 2 min-
utes each, unknown speaker number);
average-quality audio, close dialogue, i.e. where
each speaker speaks for no longer than 20 seconds
(10 files of around 10 minutes each, unknown
speaker number)
The database of known models includes 80 models
(i.e. 80 different speakers). For each person, the au-
dio file used for the test and the audio file used for
the enrollment process have been obteined in similar
condition (ambient, noise level, etc.).
The performance of a speaker recognition system is
based on FAR (“False Acceptance Rate, which mea-
sures how many speakers are falsely recognized) and
FRR (“False Rejection Rate”, which identifies the
probability that the system fails to identify a speaker
Table 1: Speaker recognition results.
High quality
< 2% 12 % 18 %
< 1% 10 % 15 %
whose model is already present in the database) (Mar-
tin et al., 1997). Table 1 shows the results.
An extensible framework for managing audiovisual
assets has been presented. A specific application of
this framework in the field of automatic audiovisual
indexing has been thoroughly described. Experimen-
tal results on a set of TV programs and audio files
have been presented, showing a quite good behavior.
Further testing will be carried out, aimed at eval-
uating the practical advantages of the proposed solu-
tion in a real-world operating environment.
Future work may take several directions: (1) de-
veloping new indexing algorithms, e.g. for recog-
nizing specific classes of objects; (2) improving the
ACTIVE framework with additional cataloging fea-
tures; (3) developing a set of general-purpose plug-
ins (e.g. for transcoding, watermarking, batch editing,
extracting embedded metadata, evaluating user pro-
files, etc.); (4) implementing a full digital asset man-
agement application on top of the ACTIVE frame-
work, in order to better exploit the work carried out
on the automatic indexing tools.
Ahonen, T., Hadid, A., and Pietik
ainen, M. (2004). Face
Recognition with Local Binary Patterns. In Proc.
ECCV, pages 469–481.
VISAPP 2016 - International Conference on Computer Vision Theory and Applications
Ajmera, J., McCowan, I., and Bourlard, H. (2004). Ro-
bust speaker change detection. IEEE Signal Process-
ing Letters, 11(8):649–651.
Barras, C., Zhu, X., Meignier, S., and Gauvain, J.-L.
(2006). Multistage speaker diarization of broadcast
news. IEEE Transactions on Audio, Speech, and Lan-
guage Processing, 14(5):1505–1512.
Berg, T. L., Berg, A. C., Edwards, J., Maire, M., White,
R., Teh, Y.-W., Learned-Miller, E., and Forsyth, D.
(2004). Names and Faces in the News. In Proc. CVPR,
pages II–848–II–854 Vol.2.
Bertini, M., Bimbo, A. D., and Pala, P. (2001). Content-
based indexing and retrieval of tv news. Pattern
Recognition Letters, 22(5):503–516.
Bradski, G. R. (1998). Real Time Face and Object Tracking
as a Component of a Perceptual User Interface. In
Proc. WACV, pages 214–219.
Bradski, G. R. (2000). The OpenCV Library. Dr. Dobb’s
Journal of Software Tools, 25(11):120, 122–125.
Chen, S. S. and Gopalakrishnan, P. S. (1998). Speaker, En-
vironment And Channel Change Detection And Clus-
tering Via The Bayesian Information Criterion. In
Proc. DARPA Broadcast News Transcription and Un-
derstanding Workshop, pages 127–132.
Dugad, R., Ratakonda, K., and Ahuja, N. (1998). Robust
Video Shot Change Detection. In Proc. MMSP, pages
El-Khoury, E., Senac, C., and Joly, P. (2010). Face-and-
clothing based people clustering in video content. In
Proc. MIR, pages 295–304.
Everingham, M. R., Sivic, J., and Zisserman, A. (2006).
“Hello! My name is... Buffy” Automatic Naming
of Characters in TV Video. In Proc. BMVC, pages
Gauvain, J.-L. and Lee, C.-H. (1994). Maximum a poste-
riori estimation for multivariate gaussian mixture ob-
servations of markov chains. IEEE Transactions on
Speech and Audio Processing, 2(2):291–298.
Khan, S., Rafibullslam, M., Faizul, M., and Doll, D. (2008).
Speaker recognition using mfcc. International Jour-
nal of Computer Science and Engineering System,
Korshunov, P. and Ooi, W. T. (2011). Video quality for
face detection, recognition, and tracking. ACM Trans-
actions on Multimedia Computing, Communications,
and Applications, 7(3):14:1–14:21.
Levenshtein, V. I. (1966). Binary codes capable of correct-
ing deletions, insertions and reversals. Cybernetics
and control theory, 10(8):707–710.
Lienhart, R., Kuranov, A., and Pisarevsky, V. (2003). Em-
pirical Analysis of Detection Cascades of Boosted
Classifiers for Rapid Object Detection. In Proc.
DAGM, pages 297–304.
Maji, S. and Bajcsy, R. (2007). Fast Unsupervised Align-
ment of Video and Text for Indexing/Names and
Faces. In Proc. MM, pages 57–64.
Martin, A., Doddington, G., Kamm, T., Ordowski, M., and
Przybocki, M. (1997). The DET curve in assess-
ment of detection task performance. In Proc. EU-
ROSPEECH, pages 1895–1898.
Meignier, S. and Merlin, T. (2010). LIUM SpkDiarization:
An Open Source Toolkit For Diarization. In Proc.
CMU SPUD Workshp.
Otsu, N. (1979). A threshold selection method from gray-
level histograms. IEEE Transactions on Systems, Man
and Cybernetics, 9(1):62–66.
Reynolds, D. A., Quatieri, T. F., and Dunn, R. B. (2000).
Speaker verification using adapted gaussian mixture
models. Digital Signal Processing, 10(1–3):19–41.
Rouvier, M., Dupuy, G., Gay, P., Khoury, E., Merlin, T.,
and Meignier, S. (2013). An Open-source State-of-
the-art Toolbox for Broadcast News Diarization. In
Sim, T., Baker, S., and Bsat, M. (2002). The CMU Pose,
Illumination, and Expression (PIE) database. In Proc.
FG, pages 46–51.
Sivic, J., Zitnick, C. L., and Szeliski, R. (2006). Finding
people in repeated shots of the same scene. In Proc.
BMVC, pages 93.1–93.10.
Smith, R. (2007). An overview of the Tesseract OCR En-
gine. In Proc. ICDAR, pages 629–633.
Viola, P. and Jones, M. J. (2001). Rapid Object Detection
using a Boosted Cascade of Simple Features. In Proc.
CVPR, pages I–511–I–518 Vol.1.
Viola, P. and Jones, M. J. (2004). Robust real-time face
detection. International Journal of Computer Vision,
Zhang, Y.-F., Xu, C., Lu, H., and Huang, Y.-M. (2009).
Character identification in feature-length films using
global face-name matching. IEEE Transactions on
Multimedia, 11(7):1276–1288.
ACTIVE, an Extensible Cataloging Platform for Automatic Indexing of Audiovisual Content