This paper describes the LIG participation to the MediaEval 2011 A ect Task on violent scenes' detection in Hollywood movies. We submitted only the required run (shot classication run) with a minimal system using only the visual information. Color, texture and SIFT descriptors were extracted from key frames. The performance of our system was below the performance of the systems using both audio and visual information but it appeared quite good in precision.
The MediaEval 2011 A ect Task: Violent Scenes
Detection is fully described in [
Our motivation was to see how a generic system for
general concept classi cationn in video shots would perform
compared to systems speci cally designed for the task like [
The descriptors were computed only on the visual information (no audio) and even only on the key frames (no motion). Three types of descriptors were used: 4
texture: a 5-scale (40-dim);
SIFT: bag of SIFT descriptors computed using Koen
van de Sande's software [
The descriptor optimization consists of two steps: power transformation: its goal is to normalize the distributions of the values, especially in the case of histogram components. It simply consists in applying an x x tranformation on all components individually. The optimal value of alpha can be optimized by crossvalidation and is often close to 0.5 for histogram-based descriptors.
PCA reduction: its goal is both to reduce the size (number of dimensions) of the descriptors and to improve performance by removing noisy components. For color and texture, the optimal number of dimension is close to half of the original one. For the SIFT-based descriptors, it is in the 150-250 range. 2.3
The classi cation was done here using a kNN-based classi er. It is a bit less e cient than an SVM one but it is much faster. 2.4
Classi cation was done separately with one kNN for each descriptor variant. The outputs of these individual classi ers are then merged at the level of normalized scores (late fusion). A linear combination of the scores is used with weight optimized on the MediaEval development set. It nally appeared that, for the MediaEval task, the SIFT descriptors did not help, compared to color and texture alone; this was not the case in the general context of TRECVID. 3.
Figure 1 shows the false alarms' rate versus miss rates for participants' best runs. It is obtained by the application of a varying threshold on the scores provided by the participants. The LIG system performs less well than other systems using both audio and visual information. However, it appears to be as good as all of them in the area of the low false alarm rates. This means that the LIG system is able to nd with a good con dence a fraction of the shots containing physical violence but beyond these, it fails to detect others, probably because the audio and/or motion modalities are necessary for them.
F-measure 0.19 0.24 0.00 0.20
MediaEval cost 8.58 6.07 10.1 7.94
While the performance of the system is consistent on Kill Bill and The Bourne Identity, it is very bad for The wizard of Oz. The system did not found any of the 46 violent shots though it predicted 60 positives (all false) in a total of 908 shots. This seems to be worse than random.
We have participated to the MediaEval 2011 a ect task with a basic system designed for general purpose concept detection in video shots. This system used only the information available in the key frames (no audio or motion). This system was initially intended to be used as a baseline and speci c extensions were considered but they could not be nalized in time. Also, concerning the target measure, the threshold was biased a bit toward recall but not enough for an optimal result with the same ranking.
In our future work, we plan to improve this baseline system by using a better classi er (SVM-based) and include motion descriptors based on optical ow and audio descriptors based on MFCC. 5.
This work was partly realized as part of the Quaero Program funded by OSEO, French State agency for innovation.