DEEP LEARNING MODELS FOR CHOREOGRAPHY GENERATION

Abstract [English]

Due to the rapid development of deep learning, the new opportunities of computational production of human movement have been opened, especially in the field of dance choreography. The paper discusses deep learning choreography generators that combine movement information, music framework, and time in order to create expressive and sensible sequences of dances. Conventional choreography models tend to have a handmade regulation or professionalized composition where flexibility and creative variety is restricted. Conversely, deep learning methods that are data driven can directly learn complex spatio-temporal patterns using big datasets of motion and video. The suggested framework uses pose representations, video frames and rhythmic information based on music to simulate the inherent relationship between motion and sound. Recurrent neural networks as LSTM and GRU models are used to learn long-lasting temporal dependencies in dance sequences whereas transformer-based models are used to improve global context awareness and sequence coherence. Also, generative adversarial networks, diffusion-based networks are explored to achieve motion synthesis which provides smooth transitions, stylistic variability and a sense of realistic continuity in movement. A modular system architecture is structured in such a way that it can allow multimodal inputs, convolutional feature extraction and temporal sequence generation. The evaluation of the experimental results is carried out on standard choreography and motion datasets and performance is measured by quantitative evaluation measures including mean absolute error, Fréchet Inception Distance, and a smoothness index specific to the movement evaluation.