Learning Neural Volumetric Representations of Dynamic Humans in Minutes