Interpretability of Correspondence Structures Under Probing

Ascertain the interpretability of correspondence-related structures encoded within the Visual Geometry Grounded Transformer (VGGT) representation space when assessed via camera-token probing of intermediate layers.

Background

The authors train simple two-layer MLP probes on VGGT’s camera tokens to recover the fundamental matrix, observing that epipolar geometry becomes decodable in the middle layers.

However, they state that the interpretability of the hypothesized correspondence structures within the representation space is not clear under the probing scheme, motivating further analysis of attention maps as a more direct lens.

References

How interpretable these structures are in the representation space is not yet clear under this probing scheme.

— On Geometric Understanding and Learned Data Priors in VGGT  (2512.11508 - Bratulić et al., 12 Dec 2025) in Section 4.1 (Does VGGT encode geometry? If so, where?)

We have not run the same test on objects other than people, so we do not claim the behaviour is specific to people: a model that matches any moving surface across frames would pass it, and for our purpose that would serve equally.

— WildHSR: Metric Feed-Forward 4D People-Scene Reconstruction from a 3D Foundation Model  (2609.29106 - Bright et al., 24 Sep 2026) in Appendix, Section “Identity Features and Pelvis Localization,” subsection “What the probe does not show”; see also Conclusion, “Limitations”

However, for disambiguation, it remains unclear which layers best separate visually ambiguous but non-covisible image pairs.

— MVDG: Efficient Multi-view 3D Disambiguation on Unconstrained Real-World Images  (2610.01098 - Xiao et al., 1 Oct 2026) in Section 3, Preliminaries, Section 3.1 context