MediaEval 2015 - UNIZA System for the "Emotion in Music" task at MediaEval 2015

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