PairBridge: Selective Optical-Flow Adaptation with Benefit–Harm Supervision
Abstract
Challenge entry: RoCo-22 (cactus8603). Robust optical-flow adaptation can improve average corruption error while introducing new errors on clean and corrupted inputs. PairBridge selects between fixed reference and adapted experts using an RGB proposal, flow verifier, and withdrawal-only guard, supervised by each switch’s endpoint error (EPE) benefit and harm. On local Spring development pairs, corruption EPE decreases by 24.11% at a 0.005% clean EPE increase; on additional pairs from seen scenes, the changes are 6.88% and 0.25%. Relative to the unguarded selector, the guard retains 98.90% and 94.75% of positive correction benefit while removing 63.46% and 52.48% of positive excess error, respectively. On additional both-frame corruptions, selection also lowers EPE by 1.50% relative to always using the adapted expert.