Abstract

This research proposes the application of four different techniques of animation storage
(Axis Angle, Quaternions, Rotation Matrices and Euler Angles), in order to determine the advantages
and disadvantages of each method through the training and evaluation of autoencoders for
reconstructing and denoising parsed data, when passing through a convolutional neural network.
The designed autoencoders provide a novel insight into the comparative performance of these animation
representation methods in an analog architecture, making them measurable in the same
conditions, and thus possible to evaluate with quantitative metrics such as Minimum Square Error
(MSE), and Root Mean Square Error (RMSE), as well as qualitatively through close observation of
the naturality, its real-time performance after being decoded in full output sequences.
My results show that the most accurate method for this purpose qualitatively is Quaternions, followed
by Rotation Matrices, Euler Angles and finally with the least accurate results:e Axis Angles.
These results persist in decoding and in simple encoding-decoding. Consistent denoising results
were achieved in the representations, up until sequences with 25% of added gaussian noise.