VGGT Knows More About Its Pose Failures Than It Reports: Linear Probe Hits 93.5% AUC
zhenjun_zhao · x · 2026-10-28
Researchers ask whether frozen VGGT internally knows when its predicted camera poses are wrong. On HardMatch—1,000 difficult image pairs where VGGT fails on 35%—a linear probe on frozen features predicts failures with 93.5% ROC-AUC, beating VGGT's own point-map confidence (79.7%) and a classifier on all its outputs (77.7%). The probe catches failures VGGT is confident about, transfers without retraining to PhotoTourism and ScanNet-1500, and emerges abruptly mid-network in the cross-view fusion layers—a pattern shared by π³.
More from Research
- Rebuilding 'Democracy Does Cause Growth': long-run GDP effect rises to 33.6% with 2024 data — ChenhaoTan · 2026-10-28
- YC-backed Wondersearch claims OpenAI-level retrieval on FiQA with no embeddings, no vector DB — antoine_chaffin · 2026-10-28
- AAAI/ICLR Reviewer: About Half of Reviews Look AI-Generated, Peer Review in Crisis — dsmirza · 2026-10-28
- A bestiary of chaotic attractors from the Field–Golubitsky map, each with Dₙ symmetry — S_Conradi · 2026-10-28
- Manifold AI's WorldScape R²SI Tops RoboDojo Agent Track by Turning Robot Failures Into Skills — jiqizhixin · 2026-10-28
- AXA Uses NVIDIA Earth-2 to Simulate 1,000+ Cyclone Seasons for Storm Risk Modeling — NVIDIAAI · 2026-10-28