Below we provide relative rotation predictions over the 3D foundation models considered in our work (VGGT, WorldMirror, and π³) over 200 randomly-sampled pairs from each of the three test sets (sELP, UnScenePairs and UnScenePairs-t).
Each page shows one image pair and its relative rotation estimates using a spherical projection visualization, with the different colors denoting the reference (black), ground-truth (blue), prediction of the pre-trained model (red), and prediction of our fine-tuned model (orange). The grey bar on top shows the model name, the overlap category and the geodesic error of both models (pre-trained and fine-tuned).
These random samples further show that our lightweight alignment procedure improves performance in extreme-view geometry settings while preserving strong performance in overlapping cases.
To navigate to dense reconstruction visualizations, please click on the "To Recon Results" button below.
Controls: