From Failure Observability to Risk-Guided Input Adaptation in Stereo Matching

Robust stereo perception requires recognizing failures under visual degradation and responding without ground-truth depth. Can frozen model-internal signals reveal failures and guide input adaptation? We study RAFT-Stereo and ACVNet across observation difficulty, within-image localization, candidate ordering, and adaptation. Their architecture-specific signals—late-iteration variability and attention entropy—track image difficulty (macro Spearman 0.866/0.886) and localize failures (within-image AUROC 0.845/0.838). Despite weak same-pixel candidate ordering, global selection among Identity and 30 fixed transformations reduces held-out six-corruption macro D1 from 15.85% to 12.89% and from 30.12% to 18.72% over three assignments, recovering about two-thirds of global-oracle headroom. A shared development-GT-selected transformation achieves comparable D1, while adaptive selection further improves 1px error. Pixel fusion leaves larger spatial-oracle gains unrealized. On CroCo-Stereo, eight-candidate selection reduces prediction disagreement by 13.3% across 20 official corruptions. Our entry achieves RbS 0.1367 and ranks second on the public Spring stereo robustness leaderboard used by the NeurIPS 2026 RoCo-Spring challenge (15 October 2026). Code for reproduction is available at https://github.com/Edw2n/stereo-risk-guided-adaptation. RoCo-Spring: Exploration Track, Stereo Estimation. Team ID: RoCo-67; Registered Team Name: Anonymous.