GMSEflow: Learning Optical Flow Through Global Matching with Squeeze-and-Excitation Networks

Estimating optical flow is crucial for numerous computer vision tasks, including autonomous vehicles, object tracking, and image frame interpolation. Yet, the challenge of improving optical flow estimation precision while managing the complexity of computational structures remains formidable. This study presents GMSEflow, an innovative model for optical flow estimation that incorporates a global multi-scale attention mechanism. We have refined the SENet architecture to allow the network to capture fine details across different scales in intricate images more effectively. The enhancements to the GMflow framework are tailored to optimize the optical flow learning process. Extensive experiments using the Sintel dataset confirm that our method outperforms other techniques in terms of accuracy. These findings represent a significant step forward in optical flow research, offering theoretical insights and practical implications for technological progress in associated domains.