Kosmo-GFM pairs a geometry foundation model with pose optimization: higher mean pose accuracy than GLUEMAP on 34 scenes, 14.07 hours instead of 69.67

Scene reconstruction no longer rests only on feature matching and geometric optimization, nor only on direct foundation-model prediction: the model supplies cameras, depth, cross-view correspondences and dynamic points, and an optimization engine turns those predictions into scene-consistent cameras and geometry. Conventional reconstruction depends on reliable visual correspondences and sufficient view overlap, so textureless regions, repetitive structures, wide baselines and scene dynamics weaken camera estimation and geometry recovery; model predictions alone do not inherently satisfy cross-view and temporal consistency across a whole scene. On the Sparse and Medium subsets of SpatialBench, Kosmo-GFM reaches absolute relative depth errors of 0.063 and 0.059, lower than the other listed methods; on the 34 scenes completed by all compared methods, PoseEngine's mean pose accuracy exceeds GLUEMAP's.

GFM learns one shared multi-view representation from which task-specific decoders read out cameras, depth, cross-view matching and dynamic points, while PoseEngine groups each center image with its neighbors into a star and obtains geometric priors and correspondences in a single forward pass. PoseEngine follows GLUEMAP's star organization with 26 images per star; unordered collections retrieve candidate neighbors with SALAD global descriptors and verify them with Doppelgangers++, while ordered sequences mainly use temporally adjacent frames. Pose evaluation covers 34 scenes and 11,336 input images, comprising 15 ETH3D, nine MipNeRF360, six Tanks & Temples and four SMERF scenes; PoseEngine reaches AUC@5 of 94.26 against GLUEMAP's 90.66.

Training data comes from KosmoWorld, which converts real-world reconstructions into image sequences with depth and camera poses for learning geometric priors. On both the Sparse and Medium subsets, adding KosmoWorld improves depth and relative-pose estimation. AbsRel falls from 0.0854 to 0.0630 on Sparse and from 0.0740 to 0.0594 on Medium; in the same comparison AUC@30 rises from 0.7197 to 0.7412 on Sparse and from 0.7107 to 0.7450 on Medium.

Evaluation covers depth, relative pose and dynamic points on the Sparse and Medium subsets of SpatialBench, plus scene-level pose and runtime on 34 scenes and 11,336 input images; 3DGS rendering scores are measured after image-conditioned refinement of test poses, not as a fixed-pose generalization test.

The scaling study does not establish an extrapolable quantitative scaling law, and gains for the 3.2B recurrent-depth variant shrink at higher training budgets, with the authors offering training-set size and diminishing returns from recurrent depth as two possible explanations; predicted 3D trajectories still show temporal jitter when the queried point is occluded in the target frame, and the effect of the static-dynamic training mixture awaits controlled ablations.