Robust Scene Flow via Task-Wise Adaptation of Pretrained Stereo and Optical Flow

RoCo-19 -- KoreaU\_DLmath. We study scene-flow estimation by adapting pretrained DEFOM-Stereo and DPFlow with their native task-specific objectives and combining their stereo disparity and optical-flow predictions through standard geometric scene-flow composition, without an additional scene-flow backbone. Shared-step comparisons favor separate task-specific supervision without an auxiliary $D2$ loss that jointly couples the stereo and flow branches, while gradient diagnostics show that the $D2$ objective is positively aligned with the native stereo loss but nearly orthogonal to the native flow loss. Robustness also depends on corruption structure: independent-view augmentation performs better when the input views are corrupted independently, whereas correspondence-consistent augmentation with ground-truth-guided transport is advantageous when corruptions preserve stereo-temporal correspondence. Our final model achieves a scene-flow 1px score of 12.200 and an SF score of 4.780 on the official RoCo-Spring evaluation. These results show that robust scene-flow adaptation depends both on how the stereo and flow experts are supervised and on how training corruptions match the structure of evaluation corruptions, within a simple framework built from pretrained stereo and flow experts. Our submission ranks first in the NeurIPS 2026 RoCo-Spring: The Robust Correspondence Challenge (Track 3: Scene Flow).