Spatially-Adaptive Instance Normalization for Generation of More Style-Recognizable Motions

Motion style transfer is a common method for generating expressive computer animation, but the motions generated by existing methods lack semantic recognition. This paper proposes a framework that generates more natural and style-recognizable motions. It utilizes a normalization layer named SP-AIN to integrate style and content features. This module generates two parameters for adaptive learning by twice convolving the style features. It retains more style features when generating new motions and improves the efficiency of style transmission. Additionally, to enhance the recognition of style features, we incorporate triplet loss as an optimization item during network training. Through a large number of examples compared with prior work, we demonstrate that our framework possesses good universality and accuracy in the task of style transfer, as well as the ability to generate high-quality stylized motions.