VGGT-Align anchors scene geometric invariants to fix scale drift in long-sequence 3D reconstruction

zhenjun_zhao · x · 2026-08-19

New arXiv paper (2608.15260, Wei Zhang et al.) tackles scale drift in chunk-based long-sequence 3D reconstruction, where the unconstrained scale DoF in sequential Sim(3) alignment lets errors compound multiplicatively, distorting global trajectories and point clouds.

Key insight: in structured environments like driving scenes, geometric quantities arising from environmental regularity are inherently invariant across temporal segments, so discrepancies in per-chunk measurements directly expose inter-chunk scale drift. The proposed Scene Geometric Invariant Anchoring (SGIA) extracts dominant geometric invariants per chunk via coarse-to-fine robust estimation and uses their cross-chunk consistency to build scale constraints independent of point cloud registration — explicitly degenerating 7-DoF Sim(3) alignment into 6-DoF rigid-body transformation and severing chain-wise scale error propagation. A lightweight test-time adaptation strategy fine-tunes only normalization-layer parameters via multi-objective self-supervision.

Original post →

More from Multimodal

Multimodal channel →