Template-Free Neural Representations for Novel View Synthesis of Humans

Traditional linear blending skinning (LBS) algorithms can act virtual avatars, but greatly rely on hand-designed body templates and 3D scan data. On the other hand, neural radiation fields can synthesis novel views from sparse views. This paper tackles the issue of obtaining accurate human performance from sparse multiviews without the use of body templates. Combining traditional LBS techniques and multilayer perceptron (MLP)-based neural representations to represent avatars proved to be very powerful in representing visual details, but the quality of the synthesized images still leaves room for worthwhile improvement. Previous research on the creation of template-free animatable volumetric actors based on neural representations has focused on the combination of the two techniques and has not suggested improvements to the internal structure of the MLP. In this paper, we present the application of self-attention to neural representations by amalgamating the benefits of transformers and neural radiation fields to establish gMLP-based neural representations.