Reconstruction of dance movements using reinforcement recurrent autoencoder and deep wavelet autoencoder

This paper proposed a new hybrid model using skeletal data to reconstruct dance movements. This proposed method combines reinforcement recurrent autoencoder (RRAE) and deep wavelet autoencoder (DWAE) to obtain temporal and spatial characteristics of dance movements. Skeletal data is extracted from high-quality dance videos through the MediaPipe Holistic algorithm and used as input to the model as a matrix. One of the critical aspects of this method is choosing the appropriate size of the input window (W) to consider the time dependence of the movements and the optimal wavelet function for it. Various experiments were conducted by altering window sizes and modifying the cost function to optimize the model's performance. This study used the RMSprop algorithm for optimization, and qualitative and quantitative evaluations were used to measure the model's effectiveness. Compared with traditional models such as RNNs, CNNs, and GANs, the proposed hybrid model has performed better regarding dynamic time warping (DTW), Euclidean distance loss (EDL), and frame-to-frame variability (FFV). Moreover, qualitative evaluations confirm that the proposed model generates natural and smooth dance movements.