Thesis Details
Generating Animations with Neural Networks
While motion capture serves as a mean for animators to circumvent some of the most arduous aspects of creating realistic animation, there is still a lot of work hiding in annotating and structuring the data. I solve this problem by designing a neural network which can be trained on a motion capture data file to reproduce human locomotion visualized in an application which allows for the user to control the character's direction. I also subject various methods of training an autoregressive model to experiments and find which method trades training times for performance the best. Additionally, I remark how the addition of certain control features to frame-by-frame generations impacts the use of recurrent neural networks for this task.
animation, motion capture, BVH, machine learning, neural networks, LSTM, discriminative models, autoregressive models
Bařina David, Ing., Ph.D. (DCGM FIT BUT), člen
Burget Radek, doc. Ing., Ph.D. (DIFS FIT BUT), člen
Češka Milan, doc. RNDr., Ph.D. (DITS FIT BUT), člen
Jaroš Jiří, doc. Ing., Ph.D. (DCSY FIT BUT), člen
@bachelorsthesis{FITBT20902, author = "Filip Dr\'{a}ber", type = "Bachelor's thesis", title = "Generating Animations with Neural Networks", school = "Brno University of Technology, Faculty of Information Technology", year = 2021, location = "Brno, CZ", language = "english", url = "https://www.fit.vut.cz/study/thesis/20902/" }