GeoCond Adapter Cuts VGGT's OOD Uncertainty Error From 0.32 to 0.20
zhenjun_zhao · x · 2026-09-17
GeoCond is a lightweight reliability adapter for frozen feed-forward 3D reconstruction backbones like VGGT, addressing silent failures under low overlap, low parallax, and extreme rotations.
- Finding: Failures are governed by geometric conditioning and poorly captured by the model's native aleatoric confidence.
- Method: GeoCond reads the backbone's predicted geometry and outputs pose-level uncertainty plus a refinement gate. Training can use frame-permutation orbit variance, ground-truth pose error, or cycle residuals from unlabelled pose graphs.
- Results: On VGGT, OOD AUSE improves from 0.32 to 0.20 (lower is better); zero-shot transfer to outdoor extreme-view scenes; avoids collapse from uniformly applying bundle adjustment. Inference needs only one backbone pass plus a small MLP.
- Extensions: Cycle-distilled variants enable ground-truth-free adaptation across backbones; the same signal supports gated refinement and pose-graph optimization.
More from Research
- "AI found new elliptic curves in days" claim disputed: it was brute-force compute on known methods — tak3sh8 · 2026-09-17
- Fixing one simple grader doubled the score—and the rest were errors in the tasks — xeophon · 2026-09-17
- A proposal for the future of scientific communication in the age of agents — tensorqt · 2026-09-17
- GPT-Policy: In-Context Robot Learning with VLM Agents, No Gradient Updates — Dongzhou Cheng · 2026-09-17
- Fathom Speeds Million-Token KV Scans 1.67x with Per-Query Read Depth — Vivek Kalyanarangan · 2026-09-17
- nnU-Net Generalization Test on BraTS-GoAT: Dice Drops from 0.906 to 0.831 Across Populations — Tristan Kirscher · 2026-09-17