Geometric relative positional encoding has become a common way to inject spatial inductive biases into multi-view transformers, according to a new arXiv preprint. The paper, titled 'HRoPE: Homography Relative Positional Encoding,' proposes a new encoding method built around homographies.
The abstract notes that existing designs condition attention on camera-level frustums or ray-marched depth. The available text cuts off at that point, so the paper's specific claims about the limitations of those approaches and the details of HRoPE itself are not included in the source. What is clear from the title is that the method uses homographies as the geometric basis for relative positional encoding.
Because the abstract is truncated, this article cannot report on experimental results or comparisons. Readers interested in the full method will need to consult the paper directly.