Information Technology Reference
In-Depth Information
The similarity measure for this descriptor is a linear combination of metrics for
histogram comparison (e.g. Bhattacharya coefficient):
x
h
sb
(
O
t
,
i
)(
x
)
h
sb
(
O
t
,
j
)(
x
)
k
sb
(
O
t
,
i
,
O
t
,
j
)=
ρ
(16)
Here
sb
stands for LL or HF and
x
is the bin index. Finally, the similarity measure
is expressed as
ρ
(
O
t
,
i
,
O
t
,
j
)=
αρ
LL
+(1
−
α
)
ρ
HF
(17)
2.3
On Scalable Content-Based Queries
A vast literature is devoted to image retrieval in large databases and much work
on video retrieval is being done. In our previous work we specifically focused on
retrieval of objects in video content [14]. In this chapter two questions in the object-
based framework: query by clip and scalable queries are addressed.
The retrieval scenario considered consists of searching for a clip in a HD video
database containing a query object. This scenario can for instance be used for detec-
tion of a fraudulent post-production, where an object is extracted from a video clip
frame by frame and inserted into the background extracted from another sequence.
Let us consider a clip
C
DB
in a video database (DB). A set of objects masks
O
DB
=
{
O
t
,
i
,
t
=
t
0
,
t
0 +
t
,...
}
is extracted for each object at each level of the
wavelet pyramid. The histogram features
H
DB
are computed and stored as metadata.
Let us then consider a query clip
C
Q
and histogram features
H
Q
of objects ex-
tracted from this clip. The user is invited to select an image
I
t
∗
∈
Δ
C
Q
in which the
object extraction result is visually the most satisfactory.
We consider both mono-level and cross-level search. In the case of mono-level
search, the descriptor
H
Q
at a given pyramid level
k
is compared to all the descriptors
available in the DB at the same level. We call this query a “mono-level” query.
Hence, a clip from the DB is the response to the query clip for a given resolution
if at least one of its frames is a “good” response to the query. The “goodness” of
a response is measured in comparison with a given threshold. This scenario is well
adapted to the scalable script in the case when the query is not transmitted with
full-resolution.
The “cross-level” search consists in comparison of a query descriptor extracted at
a chosen resolution level
k
with descriptors in DB extracted at a specified resolution
level. First of all, this type of query is interesting for a “light” processing at a client
side. The query object can be extracted on low resolution levels of wavelet pyramid
while the high resolution descriptors in the DB will be used for retrieval at server
side. Inversely, if the high-resolution descriptors are available in the original clip
(e.g. stored in the video archive), it can be compared with a low-resolution collection
of videos when searching for a fraudulent low-quality video.
In [26] main stream retrieval consisting of matching of SIFT descriptors extracted
on object masks and the global descriptor, i.e. a pair of wavelet histograms are
compared. It turns out, that firstly the HF histogram is necessary (0
<
<
1in
α
