CAR-GR for RoCo-Spring: Clean-Anchored Adaptation with Error-Supervised Gated Residual Refinement

Fine-tuning an optical-flow model on corrupted images can improve its robustness at the cost of clean accuracy. We present CAR-GR (Optical Flow Quantitative Track entry RoCo-14 · ZXUv2.0), which adapts WAFT-DAv2-A2 with ground-truth supervision on paired clean and corrupted views, an anchor that penalizes clean-output drift from the frozen parent, and a 207,139-parameter gated residual head whose gate is supervised by the error of the base prediction. Our historical challenge entry scores 0.298 Spring endpoint error (EPE) and 1.586 pixels of RobustSpring clean-to-corrupted disagreement. After we corrected a validity-mask augmentation error, three seeds lowered mean corrupted-input EPE in our local evaluation, relative to the parent, by 5.13% on development pairs and 20.66% on supplementary pairs (excluded from checkpoint selection but used in earlier screening), while mean clean EPE rose by 0.55% and 13.33%. Single-seed ablations show no benefit from the anchor or the learned gate beyond seed-to-seed variation. CAR-GR thus trades some local clean accuracy for corruption robustness; weather remains its main weakness.