mediaeval 2016 - emotional impact of movies task
TRANSCRIPT
TheMediaEval 2016Emotional ImpactofMovies Task
Run submissions• Upto5runs foreach subtask• A required run which usesnoexternal
trainingdata,only theprovided developmentdatais allowed
EvaluationMetrics:• Mean SquareError• Pearson’s Correlation Coefficient
Development dataset:LIRIS-ACCEDEDiscrete LIRIS-ACCEDE• 9800video clipsfrom 160movies underCreativeCommons licenses• Durationbetween 8sand12s• Cross-validated through acontrolled experimental protocol
Continuous LIRIS-ACCEDE• 30movies• Durationbetween 117sand4,566s(totalduration:~7hours)• Continuous induced valenceandarousal self-assessments
Testdataset:• From49newmoviesunderCreativeCommonslicenses• 1,200additional shortvideoclipsforthefirstsubtask (between8and12seconds)• 10additional longmovies(from25minutesto1hourand35minutes)forthesecondsubtask(foratotal
durationof11.48hours)
Sqdfsdf
GroundtruthValenceandarousal ranking:• Pairwise video comparisons onCrowdFlower• Annotators asked tofocusontheemotion they felt• Simpletask:• Which oneconveys themost positiveemotion?• Which oneconveys thecalmest emotion?
From rankings toratings:• Ratingscollected for40video clipsregularly distributed• 28participants• Ratingsestimated using Gaussian Process modelsContinuous annotation:• Induced valenceandarousal self-assessments• 16participants• Modified Gtrace interfaceandjoystick
Task Description• Deploy multimedia features andmodels toautomatically predict theemotional impactofmovies• Emotionconsidered interms ofinduced valenceandarousalTwo subtasks:• Globalemotion prediction:given ashortvideo clip(around 10seconds),participants’ systems areexpected to
predict ascoreofinduced valence(negative-positive)andinduced arousal (calm-excited)forthewhole clip;• Continuous emotion prediction:asanemotion felt during ascene may be influenced bytheemotions felt during
theprevious ones,thepurpose here is toconsider longervideos,andtopredict thevalenceandarousalcontinuously along thevideo.Thus,ascoreofinduced valenceandarousal should be provided foreach 1s-segmentofthevideo.
Context• Anevolution ofprevious years tasks onviolenceandaffectprediction from videos• Applications:• Personalized contentdelivery• Movie recommendation• Video editing supervision• Video summarization• Protectionofchildren from potential harmful content
Organizers:EmmanuelDellandréa,Liming Chen,YoannBaveye,MatsSjöberg,ChristelChamaretContact:EmmanuelDellandréa – [email protected]
Representation ofemotions
Creditsandlicenseinformationisavailablehere:
http://liris-accede.ec-lyon.fr/database.php
Arousal Valence