mood-based classification of tv programmes - jana eggink, sam davies, denise bland (semantic media @...
DESCRIPTION
This talk was given by Jana Eggink, Sam Davies and Denise Bland (BBC R&D) at the "Semantic Media @ BBC" event on 6 February 2013.TRANSCRIPT
R&D BBC MMXIII
Mood-based Classification of TV Programmes
Jana Eggink, Sam Davies, Denise Bland
BBC R&D
{jana.eggink, sam.davies, denise.bland}@bbc.co.uk
R&D BBC MMXIII
Searching the Archives
• BBC aims to open up its archives for public access by 2022
• Limited metadata available
• Title
• Broadcast date
• Genre (mostly)
• Limited: actors, semantic labels for professional use
• Mood as additional metadata, intuitive understanding
R&D BBC MMXIII
ActivityPotency
Which Moods?
Evaluation
(EPA model based on Osgood et al, 1957)
Interesting – Boring
Happy – Sad
Light-hearted – Dark
Serious – Humorous Fast paced – Slow paced
Exciting – Relaxing
R&D BBC MMXIII
User Trial
• 200 members of the general public
• 544 video clips (3 minutes excerpts)
• Each labelled by at least 6 participants
R&D BBC MMXIII
Inter-rater Agreement
Krippendorff’s Alphae
o
D
D1
Agreement about Mood Labels
random
perfect
R&D BBC MMXIII
Correlation
• Which moods are independent?
• Do observed correlations correspond to the EPA model?
happy humorous exciting interest fast light
happy
humorous
exciting
interest
fast
light
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
-0.6 -0.4 -0.2 0 0.2 0.4 0.6
-0.6
-0.4
-0.2
0
0.2
0.4
0.6
happyhumorous
excitinginterest fast
light-hearted
sadserious
relaxingboringslow
dark
Component 1
Com
pone
nt 2
PCACorrelation
63% variance
24%
var
ianc
e
R&D BBC MMXIII
Video clips
• 444 in development set, 3-fold cross validation
• 100 in holdout set
Features
• Audio (MFCCs, amplitude, zero-crossing, spectral centroid and roll-off)
• Video (face, luminance, cuts, motion)
• Genre (human assigned)
Testing
• Clips with very clear moods only
• Average rates, all clips on a 1 to 5 scale
Classification
R&D BBC MMXIII
Automatic Classification Gives Good Results
Clear moods only
• 2 class problem
• >95% correct for serious/humorous
• ~90% correct for slow/fast-paced
Classification Accuracy
R&D BBC MMXIII
Automatic Classification Gives Good Results
Average rates
• 1-5 scale
• ~0.7 RMSE for serious/humorous
• <0.7 RMSE for slow/fast-paced
RMS Error for detailed moods
R&D BBC MMXIII
Conclusions
• There is general agreement about mood for TV programme clips
• Mood perception is dominated by two dimensions
• Classification for clips with clear moods is very accurate, and still possible on a detailed continuous scale
• Both genre labels and signal processing features are useful
• Humorous-serious is strongly related to genre
• Slow/fast-paced can be better modelled by audio/video features
Eggink & Bland, A Large Scale Experiment for Mood-Based Classification of TV Programmes, IEEE Int. Conf. Multimedia and Expo, ICME2012, also as BBC White Paper Nr. 232
R&D BBC MMXIII
Demo
R&D BBC MMXIII
• Usage data 14th May 2012 to 22nd August 2012
• 3206 unique users, nearly a third (1013) are returning users
Usage of the Redux Mood GUI
R&D BBC MMXIII
Search Behaviour
R&D BBC MMXIII
Frequent Programmes Watched
Never Mind the Buzzcocks 258
Torchwood 90
Dr Finlay`s Casebook 74
An Evening in with David Attenborough 55
Holiday Weatherview 49
Would I Lie to You? 46
Never Mind the Buzzcocks 36
Morecambe and Wise 33
Never Mind the Buzzcocks 32
Till Death Us Do Part 32
Programmes Watched
R&D BBC MMXIII
Outliers attract Attention
R&D BBC MMXIII
• Public facing Mood GUI based on iPlayer
• Available; http://moods.ch.bbc.co.uk
• Requires greater research in UX
Outlook and Future Work
R&D BBC MMXIII
• Integration of pre-existing metadata
Outlook and future work