Exploiting Point Motion, Shape Deformation, and Semantic Priors for Dynamic 3D Reconstruction in the Wild