Linear probe predicts VGGT pose failures the model does not report
Original title: Does VGGT Know What It Does Not Know?
On HardMatch, where VGGT fails on 35% of 1,000 pairs, a frozen-feature probe reaches 93.5% ROC-AUC.
VGGT predicts cameras and geometry in one pass but reports no uncertainty for relative poses. On HardMatch, a linear probe of frozen features predicts pose failure on held-out categories with ROC-AUC 93.5%, above VGGT's point-map confidence at 79.7% and a classifier on all outputs at 77.7%. The authors say the signal arises during two-view fusion, transfers to rotation failures on PhotoTourism and ScanNet-1500, and also appears in pi^3.
Unconfirmed: Probe performance is reported by the paper's authors on their benchmarks.
- The probe catches failures that VGGT itself treats as confident.
- It can let a downstream system abstain on distrusted pairs.
- The per-image encoder does not contain this failure signal, the authors report.
From the original article
1 min read at arxiv.orgFeed-forward 3D reconstruction models such as VGGT regress cameras and geometry in a single forward pass, but they report no uncertainty for the camera poses they predict: an incorrect relative pose looks no different from a correct one. We ask whether the frozen model nonetheless knows when its pose is wrong.
By Adityan Thirumani Srinivasan · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.