Less Decoder is More Encoder: Geometric Representation Learning from Novel View Synthesis
SNAP, a self-supervised pose-conditioned transformer for novel view synthesis, learns transferable geometric representations competitive with geometry-supervised methods across five tasks.