Search

Physically based cloth simulation requires a model that represents cloth as a collection of nodes connected by different types of constraints. In this paper, we present a coefficient prediction framework using a Deep Learning (DL) technique to enhance video summarization for such simulations. Our proposed model represents virtual cloth as interconnected nodes that are subject to various constraints. To ensure temporal consistency, we train the video coefficient prediction using Gated Recurrent Unit (GRU), Long-Short Term Memory (LSTM), and Transformer models. Our lightweight video coefficient network combines Convolutional Neural Networks (CNN) and a Transformer to capture both local and global contexts, thus enabling highly efficient prediction of keyframe importance scores for short-length videos. We evaluated our proposed model and found that it achieved an average accuracy of 99.01%. Specifically, the accuracy for the coefficient prediction of GRU was 20%, while LSTM achieved an accuracy of 59%. Our methodology leverages various cloth simulations that utilize a mass-spring model to generate datasets representing cloth movement, thus allowing for the accurate prediction of the coefficients for virtual cloth within physically based simulations. By taking specific material parameters as input, our model successfully outputs a comprehensive set of geometric and physical properties for each cloth instance. This innovative approach seamlessly integrates DL techniques with physically based simulations, and it therefore has a high potential for use in modeling complex systems.