Establish whether the pretrained correspondence signal improves end-to-end tracking

Establish whether the correspondence signal in the pretrained VGGT-$\Omega$ query-key representation improves the accuracy of the complete end-to-end WildHSR tracking pipeline, rather than only improving association when upstream person proposals and bodies are fixed.

Background

The paper demonstrates that intermediate VGGT-Ω\Omega features can retrieve the same moving person across frames and reports a controlled association experiment combining projected identity descriptors with motion and confidence cues. However, that component experiment holds the Multi-HMR proposals and bodies fixed and therefore does not isolate the effect of the correspondence representation in the complete reconstruction system.

The authors explicitly state that the probe and component experiment do not establish an end-to-end tracking benefit. A remaining problem is therefore to measure and attribute the contribution of the pretrained correspondence signal when proposal generation, body reconstruction, window composition, and association all operate jointly.

References

The probe alone also cannot establish that its correspondence improves tracking. The association-component experiment in Table~\ref{tab:identity_assoc} tests how descriptor and motion cues combine when upstream proposals and bodies are held fixed; it does not isolate their effect in the final end-to-end pipeline.

— 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”