Detecting Violent Content in Hollywood Movies and User-Generated Videos - Smart Information Systems: Computational Intelligence for Real-Life Applications - page 302

11.5 Application
We present in this section a browser-based visualization tool that allows users to
explore movies and online videos based on the detected violence levels. In this tool,
currently, only the objective violence definition of the MediaEval VSD task is used to
model violence. The system offers the visualization of annotations and results of the
MediaEval 2013 VSD task [ 9 ] and can interactively download and analyze content
from video hosting sites such as YouTube.
The development and evaluation of VSD creates the need for a detailed visu-
alization to assess the strengths and weaknesses of algorithms. Our visualization
tool [ 27 ] consists of three parts: the Ranked List view shows the results on the test set
of the MediaEval 2013 VSD task, the Annotations view shows the annotations of the
MediaEval 2013 VSD training set and the Online Analysis carries out our analysis
pipeline [ 1 ] to arbitrary online videos.
11.5.1 The Method
Among the plurality of audio features, MFCCs are shown to be indicators of the
excitement level of video segments [ 36 ]. Therefore, we employ them as low-level
audio features. For the representation of video segments, we use mid-level audio fea-
tures based on MFCCs in a BoAW scheme. We apply the BoAW approach with two
different coding schemes; as an alternative to sparse coding (introduced in Sect. 11.3 ),
we also carried out vector quantization. We train a pair of two-class SVMs in order
to learn violence models using both mid-level feature representations. Normally, in a
basic SVM, only class labels or scores are output. The class label results from thresh-
olding the score, which is not necessarily a probability measure. The scores output by
the SVM are converted into probability estimates using the method explained in [ 34 ].
11.5.2 Ranked List
The user first selects the run (algorithm and parameters), of which the results will
be visualized. The user can also select a specific test movie or the whole test set.
The Ranked List view (Fig. 11.6 ) then shows the thumbnails of all segments with an
overlay of the violence score (i.e., the probability of violence), time information and
a notice whether the classification matches the ground truth. If a segment is classified
as violent, the thumbnail is highlighted with an orange frame around it. This enables
the user to interpret the results easily and quickly. A click on the thumbnail plays
the given segment without leaving the Ranked List view. The user can sort the list
by the violence scores returned by the algorithm, or can sort it by time to see the
classification results chronologically from the beginning to the end of the movie.
We also added a button which, when pressed, jumps to a random part of the list to
enable a more dynamic exploration experience.