Title: FlowIt: Achieving Global Matching through Hierarchical Transformers and Optimal Transport for Optical Flow

Abstract: This paper introduces FlowIt, an innovative architecture designed for optical flow estimation that integrates global matching mechanisms with refinement processes guided by confidence and occlusion data. Central to the FlowIt design is a hierarchical transformer framework capable of capturing broad global context, which allows the model to accurately represent long-range correspondences. To address the constraints inherent in localized matching strategies, we define the initial flow estimation as an optimal transport problem. This approach generates a highly resilient initial flow field while simultaneously producing explicit occlusion and confidence maps. These indicators are then incorporated into a guided refinement phase, during which the network systematically transfers reliable motion estimates from areas of high confidence to regions characterized by ambiguity and low confidence. We validated our method through comprehensive experiments on the Sintel, KITTI, Spring, and LayeredFlow datasets. Our results demonstrate that FlowIt sets a new state-of-the-art on the competitive Sintel benchmark. Furthermore, it achieves superior cross-dataset zero-shot generalization performance on Sintel, Spring, and LayeredFlow, while maintaining competitive results in both standard and zero-shot generalization settings on the KITTI benchmark.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC