Adaptive Geodesic Conformal Prediction for Egocentric Camera Pose Estimation

Egocentric pose estimation for AR and assistive devices demands guaranteed uncertainty regions, not just accurate predictions. Conformal prediction (CP) provides such guarantees without retraining, but standard CP with a single fixed threshold achieves nominal 90% overall coverage while covering only ~60% of the hardest 25% of frames (Q4) — a ~30 percentage-point conditional coverage gap consistent across 12 participants, 3 predictors, and 3 horizons (108 evaluations) on EPIC-Fields. The researchers also show that a geodesic SE(3) nonconformity score identifies physically harder frames than Euclidean scoring, with only 15-26% Q4 overlap and 2-3x higher ground-truth camera displacement.

To close this gap, the team proposes DINOv2-Bridge adaptive CP: a two-stage difficulty estimator trained on a single source participant that transfers cross-participant without any images at test time. This method improves Q4 coverage from ~0.75 to ~0.93 while maintaining overall coverage at the 90% target. The approach is particularly significant for real-world AR applications where difficult frames (e.g., rapid head motion, occlusions) are common, and reliable uncertainty quantification is critical for user safety and experience. The work underscores the limitations of one-size-fits-all conformal prediction and offers a practical, training-free solution for adaptive uncertainty sets.