CroCo-StableBridge: Risk-Aware Input Routing for a Frozen Stereo Matcher
Abstract
Challenge entry: RoCo-22 · cactus8603. Simple preprocessing can improve stereo matching under image corruption, but a treatment that helps one pair may harm another. We present StableBridge, an input selector for a frozen CroCo-Stereo matcher. Early encoder features propose denoising or sharpening and estimate its benefit; deeper features estimate its harm. StableBridge applies the proposal only when the benefit score exceeds the harm score by a fixed margin and otherwise leaves the pair unchanged. In scene-held-out cross-validation on Spring with synthesized corruptions, it reduces corrupted-crop mean absolute dis parity error (Abs) by 5.08% without changing clean-crop error. At matched total modification counts over clean and corrupted examples, the harm check lowers the fraction of changed corrupted crops that worsen by 4.55 percentage points, although it does not clearly improve overall gain under scene resampling. The official system obtains RbS 0.1456; without a matched CroCo-only submission, this is an end-to-end result rather than an estimate of the selector’s incremental effect.