SPEAR: Staged Pre-training and Expert Adapter Routing for Generalizable Optical Flow Learning

As optical flow estimation matures, accurately recovering motion at ever higher resolutions and finer detail has become an increasingly important goal. The RoCo-Spring benchmark, hosted as part of the NeurIPS 2026 Competition Track, was established to push this frontier: beyond a clean, general domain, it demands robust estimation under a wide range of image corruptions. To address this challenge, we present SPEAR, a robust optical flow model built from staged pre-training and image-conditioned residual adaptation. In the first stage, we build on WAFT and adopt a larger, richer feature extractor, starting from the pre-training of its projection layers to strengthen flow accuracy in the clean domain. In the second stage, for robustness adaptation, we introduce a router together with expert adapter modules that provide corruption-specific corrections according to the type of degradation. The router and adapters are trained while the clean-domain base estimator is kept frozen, so that robustness is acquired while largely preserving clean-domain accuracy. This staged approach enables SPEAR, developed by Team RoCo-23, to achieve state-of-the-art performance on the official Spring benchmark across both standard and challenging conditions. Code and other resources are available at \url{https://final-0.github.io/SPEAR/}.