Information Technology Reference
In-Depth Information
Fig. 11.6
Ranked List
view of our visualization tool
11.5.3 Annotations
The training set of the MediaEval 2013 VSD task provides annotations for 18 Hol-
lywood movies. The annotations mark the presence of audio, visual and audio-
visual concepts such as explosions, gunshots, screams, blood, fights, car chases, fire,
firearms, coldweapons and gore. The user can query any or all movies for any of these
concepts (e.g., show all segments with fire in
Saving Private Ryan
). The annotations
are then displayed in a view (Fig.
11.7
) similar to the one of the
Ranked List
.
11.5.4 Online Analysis
The
Online Analysis
(Fig.
11.8
) executes our VSD pipeline to any video hosted by
YouTube (or any other site supported by the youtube-dl script). After the user entered
the URL, the video is downloaded, transcoded, and split into segments. The MFCC
feature vectors of the audio of each segment are subsequently computed and used
to build mid-level features with sparse coding and vector quantization as explained
in [
1
]. Both mid-level feature representations are used to classify the segment and
produce two violence scores. Even though our methods only use audio features, the
Online Analysis
pipeline can be applied to any method using audio, visual, or audio-
visual features. In addition to the
Ranked List
view, the
Online Analysis
produces a
Search WWH ::
Custom Search