Describir: Differentiable 3D Scene Representations With Point-Based Neural Methods