Vision foundation models can predict depth, pose, and point clouds in a single forward pass, yet they leave multi‑view geometry unchecked. Self‑Geometry shows that enforcing explicit epipolar consistency at inference tightens those predictions across diverse datasets without requiring full model retraining.
Prior test‑time methods rely on implicit self‑consistency derived from a model’s own outputs, which delivers only limited gains when the pretrained VFM is already inaccurate. The new pipeline replaces this weak signal with pseudo ground‑truth 2D correspondences and optimizes them directly against multi‑view and epipolar losses.
On the wide‑baseline ETH3D benchmark, Self‑Geometry lifts pose accuracy of VGGT by 9.2 % (AUC@30) and 37.3 % (AUC@3), while a comparable model gains 5.0 % and 25.1 % respectively, demonstrating that epipolar constraints translate into sizable improvements on challenging scenes [1]. Across six VFMs—VGGT, π³, DA3‑Giant/Large/Base/Small—and four standard suites (7Scenes, ETH3D, ScanNet++, HiRoom), the same adaptation consistently raises both pose AUC and depth F1 scores.
The approach still depends on reliable 2D correspondences; in texture‑poor or dynamic environments those matches can be noisy or missing. Moreover, although the LoRA‑based lightweight TTA runs in under two minutes per scene on an RTX PRO 6000, it remains a non‑trivial preprocessing step that precludes strict real‑time deployment.
If these gains hold broadly, future pose‑and‑depth pipelines should treat test‑time geometric adaptation as a default plug‑in rather than an optional afterthought, and benchmark suites ought to include a “Self‑Geometry” baseline when reporting VFM performance.