Novel-View Human Action Synthesis

Novel-View Human Action Synthesis aims to synthesize the movement of a body from virtual viewpoint, given video real viewpoint. We present novel 3D reasoning target first estimate mesh and transfer rough textures 2D images mesh. As this may generate sparse on due frame resolution or occlusions. produce semi-dense textured by propagating transferred both locally, within local geodesic neighborhoods, globally, across symmetric semantic parts. Next, we introduce context-based generator learn how correct complete residual appearance information. This allows network independently focus learning foreground background synthesis tasks. validate proposed solution public NTU RGB+D dataset. The code resources are available at https://bit.ly/36u3h4K.