RayRoPE: Projective Ray Positional Encoding for Multi-View Attention

RayRoPE is a novel positional encoding technique for multi-view transformer architectures. The method represents patch positions using rays and predicted 3D points, enabling SE(3)-invariant attention with multi-frequency similarity. This approach gives AI developers a geometrically adaptive way to process posed input images in 3D vision tasks.

Opening Kapyn…