Less Decoder is More Encoder: Geometric Representation Learning from Novel View Synthesis
arXiv cs.AI · 3 Oct 2026, 03:59 AEST
This paper examines the role of Novel View Synthesis (NVS) in geometric representation learning. It presents SNAP, a self-supervised encoder-decoder transformer that improves representation capabilities in NVS tasks.
Key facts
- In principle, NVS should reason about 3D scene structure, thereby enabling transferable multi-view geometric representations.
- SNAP is task agnostic, and we show that it is competitive with special-purpose geometry-supervised methods.
- SNAP's patch features exhibit emergent viewpoint invariance that approaches heavily supervised models despite lower compute and data budgets.
Read aloud
This paper examines the role of Novel View Synthesis NVS in geometric representation learning. It presents SNAP, a self-supervised encoder-decoder transformer that improves representation capabilities in NVS tasks. In principle, NVS should reason about 3D scene structure, thereby enabling transferable multi-view geometric representations. SNAP is task agnostic, and we show that it is competitive with special-purpose geometry-supervised methods. SNAP's patch features exhibit emergent viewpoint invariance that approaches heavily supervised models despite lower compute and data budgets.
Sources in this cluster
- Less Decoder is More Encoder: Geometric Representation Learning from Novel View Synthesis
arXiv cs.AI · 3 Oct 2026, 03:59 AEST