MediaEval 2015 - UNIZA System for the "Emotion in Music" task at MediaEval 2015
Post on 15-Apr-2017
76 Views
Preview:
Transcript
UNIZA System for the „Emotion in Music“ task
at MediaEval 2015
Ing. Miroslav Malík Department of Telecommunica1ons and Mul1media Faculty of Electrical Engineering University of Žilina, Slovakia
UNIZA system
1
Par9cle Swamp Op9miza9on / Gene9c Algorithm
• Parallel run of PSO / GA
• Best individuals from both algorithms selected into next itera1on of the op1miza1on process
• fitness func1on -‐ RMSE between predicted and ground truth labels
2
Audio Features
• Op1mal features from baseline set • Features from MIR Toolbox
• 2 best combina1ons used: – Op#mal_1 – 139 features – Op#mal_2 – 129 features – 72 iden1cal features (MFCC, slope, skewness...)
3
Results of our system on the development data
4
5
Official results
• (almost) equal result obtained using only approximately 50% of the baseline features
• it might be useful to apply the SVR with different kernel parameters and/or also different feature sets for each emo1onal dimension
6
Conclusion and future work
THANK YOU FOR YOUR ATTENTION
top related