The authors identify that standard RoPE encodes relative positions in per-image coordinates, which is insufficient for cross-view attention because a 3D point has different coordinates in different views.
WorldRoPE encodes patch tokens as 3D Gaussians to ground multi-view attention
Kim, Lim, and Han show that integrating rotary embeddings over projected 3D token supports improves multi-view Transformers on novel view synthesis, depth estimation, and 3D reconstruction.
Top University
Mijeong Kim · Kyujin Lim · Bohyung Han
Seoul National University
Research Digest··3 min read
Mijeong Kim, Kyujin Lim, and Bohyung Han introduce WorldRoPE, a positional encoding that models each image patch as volumetric Gaussian primitives in a shared world frame.
Why this paper
From Seoul National University
In one line
Modeling token positions as Gaussian primitives in a shared world frame and integrating rotary attention over their uncertainty improves multi-view Transformers.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ·No stated limitations found
- ·No benchmark numbers found
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§