Robust Stereo Matching under Missing Correspondence Evidence

RoCo-19 -- KoreaU_DLmath. Stereo matching fundamentally relies on reliable cross-view correspondence, yet common robustness strategies primarily perturb image appearance without necessarily disrupting the binocular evidence used for matching. We identify a distinct failure mode in which corruption removes valid correspondence cues or introduces spurious ones, causing large disparity drift. Our analysis of CroCo-Stereo on Spring/RobustSpring shows that such corruption-induced instability is strongly concentrated in evidence-poor regions, including sky and low-texture areas. Motivated by this observation, we train explicitly under degraded correspondence evidence. Our framework combines geometry-aware and sensor-locked image-space corruption with clean-corrupted prediction consistency and structured epipolar correspondence dropout. The proposed dropout masks the epipolar key band containing the expected cross-view match for selected decoder queries, directly training the model when local binocular correspondence is unavailable. A frozen clean teacher preserves the clean stereo solution, while a lightweight monocular dense prediction transformer branch provides complementary geometric cues through conservative fusion. All corruption and correspondence-dropout operations are restricted to training, and inference uses a single fixed checkpoint without test-time adaptation or corruption-specific processing. On the official RobustSpring stereo evaluation, our submission reduces corruption-induced $\Delta\mathrm{Abs}$ from $2.386$ to $1.441$ (39.6%), $\Delta D1$ from $4.427$ to $1.909$ (56.9%), and the robustness-balanced Spring score (RbS) from $0.1456$ to $0.1176$ (19.2%) relative to the CroCo reference, while improving clean Abs from $0.471$ to $0.457$. These results rank our submission first in the NeurIPS 2026 RoCo-Spring: The Robust Correspondence Challenge (Track 2: Stereo Matching) and support explicit training for correspondence-evidence failure as an effective strategy for robust stereo matching.