Reproduce “WorldRoPE: Probabilistic World-Centric Rotary Embedding for Multi-View Attention” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Mijeong Kim, Kyujin Lim, Bohyung Han
WorldRoPE represents each image patch as multiple possible three-dimensional regions, capturing depth uncertainty and patch extent. It projects these regions into another camera and averages positional rotations in attention, attenuating contributions according to uncertainty. In LVSM on RE10K, it achieves 26.75 dB PSNR, versus 26.26 dB for URoPE, under the same backbone and training protocol.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.