In this work we predict the emotional impact of videos using the LIRIS ACCEDE dataset. By feeding the sequen- tial video features into an Echo State Network, we tackle the problem of varying video length while taking temporal history of the video into account. First, Mean-Pooling is evaluated as a baseline, against which the performance of the fine-tuned Echo State Network is evaluated. Other than expected, this approach could not provide any improvements in prediction performance compared to our baseline, and prior work as well.


  • No labels