MoGe is a powerful model for recovering 3D geometry from monocular open-domain images. The model consists of a ViT encoder and a convolutional decoder. It directly predicts an affine-invariant point ...