Learning to Regress Bodies from Images using Differentiable Semantic Rendering