GeoCond: A Conditioning-Aware Reliability Adapter for Feed-Forward 3D Reconstruction
David Ahmedt-Aristizabal, Mohammad Ali Armin, Russell Tsuchida, Lars Petersson
cs.CV
2026-09-16
CSIRO attaches a tiny reliability head to frozen VGGT, predicting pose uncertainty from geometric conditioning and cutting OOD AUSE from 0.32 to 0.20.
Feed-forward 3D models such as VGGT predict cameras, depth, and point maps in one pass, which is why they are attractive as SfM front-ends. They can fail quietly. Under low overlap, low parallax, or near-180° relative rotation, a relative pose can be off by tens of degrees while per-pixel confidence still looks high.
Uniform bundle adjustment (BA) is not a fix. BA helps pairs with enough track support and can wreck low-overlap pairs whose geometric constraints are weak. What is missing is a pose-level trust signal: keep this relative pose, refine it, drop it, or capture another view.
GeoCond is a small MLP, about 19.6k parameters, attached to a frozen backbone. It does not update VGGT. It reads geometry the backbone already predicted and packs seven scalars: predicted parallax, co-visibility, baseline, relative-rotation magnitude, the camera head's rotation and translation corrections, and negated mean native confidence. Concatenated with the relative pose as a quaternion plus unit translation, the head outputs pose uncertainty u, a refinement gate g, and an optional SE(3) residual.
The backbone stays frozen. Supervision can come from three teachers. The default VGGT recipe shuffles the same image set K=6 times; if the relative translation direction moves with the permutation, the pair is unstable (orbit variance). When labels exist, the target is pose error. When they do not, independent pairwise pose graphs supply cycle residuals: Rij Rjk Rki should be identity. That last route matters for permutation-equivariant models such as π³, where orbit variance collapses to zero.
At inference the default head needs one backbone pass and the MLP. Ten pairs cost 0.17 ms, about 0.2% of a 95.6 ms backbone pass.
The main setting freezes VGGT-1B. Indoor tests use 7-Scenes; the outdoor extreme-view set is MegaUnScene. On 7-Scenes heads, median pose error rises from 3.0° at high parallax to 17.4° at low parallax. On MegaUnScene it rises from 3.3° to 44.2°. In the office scene, pairs with overlap below 0.01 have median error 42.2° and a 54% catastrophic rate above 30°.
| Setting | Metric | GeoCond | Baseline |
| Scene-disjoint OOD | AUSE↓ | 0.20 | native confidence 0.32 |
| Indoor | AUSE↓ | 0.17 | native 0.19; K=6 teacher 0.17 |
| Indoor gated AUC@30↑ | 53.5 | feed-forward 51.6; BA always 53.6 | |
| OOD gated AUC@30↑ | 58.4 | feed-forward 60.3; BA always 38.4 |
Out of distribution, the gate's job is to block harmful refinement, not to beat feed-forward on every pair. Dropping the camera-head correction terms raises OOD AUSE from 0.209 to 0.273; keeping only the four basic geometric scalars collapses it to 0.359. A five-member deep ensemble and MC-dropout reach about 0.20 OOD AUSE as well, at roughly five backbone passes.
The same score is usable downstream. Uncertainty-weighted rotation averaging cuts MegaUnScene median global rotation error from 15.1° with uniform weights to 12.7°. Adaptive capture matches all-8-frame accuracy with about 3.0 frames versus 3.85 for native confidence. At a catastrophic-label rate α=0.10, GeoCond keeps 64% of pose pseudo-labels versus 36% for native confidence.
Feed-forward 3D is already fast enough to be a module. It still needs a trust interface. GeoCond does not claim a better pose on every pair. It answers whether the pose should be trusted. For SLAM front-ends, pose-graph fusion, and capture planning, that is a drop-in layer. The backbone stays frozen and inference stays single-pass.
The authors are direct about this. The K-pass orbit teacher remains stronger when its compute budget is available. Absolute failure-probability calibration is domain-specific; ranking transfers more cleanly than the probability scale. If native confidence already tracks the dominant failure mode, target-domain supervision or recalibration may still be required. Evaluation stops at relative-pose AUC, uncertainty ranking, retained-pair quality, and rotation-averaging front-ends. Full SfM/SLAM trajectory error is left for later. OOD gated AUC@30 is 58.4, still below feed-forward-always at 60.3. In extreme views the refinement gate is a fuse, not an accelerator.