Information Technology Reference
In-Depth Information
- spatio-temporal partitioning, that is extraction of meaningful objects from video
in order to detect events of interest related to objects or to realize object-based
queries on video databases [14].
In the first case, the key feature to be extracted from video is represented by shot
boundaries which are considered as a “mid-level” semantic feature compared to
high level concepts appealing to human-like interpretation of video content. A shot
is a sequence of frames corresponding to one take of the camera. Very vast literature
has been devoted to the problem of shot boundary detection in the past. This was
the subject of TRECVid competition [15] in 2001-2007 and various industrial solu-
tions have come from this intensive research. Shot boundaries are the most natural
metadata which can be automatically generated and allow for a sequential naviga-
tion in a coded stream. Scene boundaries correspond to the changes of the content
with more semantic interpretation, they delimit groups of subsequent shots which
convey the same editing ideas. In the framework of scalable indexing of HD video
we wish to analyze a new trend for efficient video services: embedded indexing in
Scalable Video coding. This is namely one of the focuses of JPSearch initiative [16]:
embedding of metadata into the data encoded in the JPEG2000 standard. Hence the
latest research works link content encoding and indexing in the same framework be
it images or videos [3].
2.2.1
Embedded Mid Level Indexing in a Scalable Video Coding
The coding scheme [3] inspired by JPEG2000 scalable architecture allows a joint
encoding of a video stream and its temporal indexing such as shot boundaries in-
formation and key pictures at various quality levels. The proposed architecture of
scalable code-stream consists of a sequences of Groups of Pictures (GOPs) corre-
sponding to a video shot and containing its proper source model which at the same
time represents a content descriptor.
The video stream architecture is depicted in Figure 4. The authors propose using
Vector Quantization (VQ) to build a visual code-book
VC
for the first key-picture
of each shot. Here the encoding follows a traditional scheme in VQ coding: the
key-picture is split into blocks forming vectors in the description space. Then an
accelerated K-means algorithm is applied resulting in a VC of pre-defined dimen-
sion. The key-picture is then encoded with this code-book and the error of vector
quantization is encoded using a JPEG2000 encoder. For all following key-pictures,
the code-book obtained is applied for their encoding. If the coding distortion for a
given key-picture
I
ij
in the shot
S
i
encoded with
VC
i
is higher than a pre-defined
Fig. 4
Temporal decompo-
sition of a video stream
Key-
picture
Key-
picture
Key-
picture
B
1
picture
B
0
picture
B
1
picture
…
GOP
n
GOP
i-1
GOP
i
