A new arXiv preprint introduces ReDT, a feed-forward visual geometry transformer built to scale multi-view reconstruction to high-definition inputs. The authors position it against existing visual geometry transformers, which are typically trained at a fixed, standard definition and therefore struggle to generalise to higher resolutions.
The key idea, as the name suggests, is a resolution-decoupled design. Rather than tying the transformer's operation to a specific input resolution, ReDT appears to separate the resolution handling from the geometry reasoning, allowing the same model to work on HD data. The abstract does not provide architectural details or quantitative results, so the claims are limited to the proposed approach.
As a preprint, this work has not yet been peer-reviewed. The main contribution, according to the abstract, is the conceptual shift toward resolution-decoupled training for multi-view geometry, which could be useful for applications requiring fine detail, such as 3D reconstruction from high-resolution images.