GDL

Creating dance animations can be done manually or using a motion capture system. An intelligent system that able to generate a variety of dance movements should be helpful for this task. The recurrent neural network such as Long Short-Term Memory (LSTM) or Gated Recurrent Unit (GRU) could be trained as a generative model. This model is able to memorize the training data set and reiterate its memory as the output with arbitrary length. This ability makes the model feasible for generating dance animation. Remo is a dance that comprises several repeating basic moves. A generative model with Remo moves as training data set should make the animation creating process for this dance simpler. Because the generative model for this kind of problem involves a probabilistic function in form of Mixture Density Models (MDN), the random effects of that function also affect the model performance. This paper uses LSTM and GRU as generative models for Remo dance moves and tests their performance. SGD, Adagrad, and Adam are also used as optimization algorithms and drop-out is used as the regulator to find out how these algorithms affect the training process. The experiment results show that LSTM outperforms GRU in term of the number of successful training. The trained models are able to create unlimited dance moves animation. The quality of the animations is assessed by using visual and dynamic time warping (DTW) method. The DTW method shows that on average, GRU results have 116% greater variance than LSTM’s.

Beri Komentar ?#(0) | Bookmark