HRoPE for Surface-Aware Positional Encoding in Multi-View Transformers
October 19, 2026
HRoPE introduces Homography Relative Positional Encoding to ground patch tokens in physical surface geometry rather than sensor-centric views. It represents local geometry as slanted 3D planes to improve spatial inductive biases in multi-view transformers.
HOW THIS AFFECTS YOU
●
researcherThis method improves how multi-view models handle scene geometry by moving beyond ray-marched depth or frustum-based encodings.