Enhanced Optical Flow Estimation via Multiscale Kernel Selection and Super-Resolution Integration

Optical flow estimation, pivotal in computer vision, captures motion between adjacent frames. Traditional methods like RAFT suffer from limited feature capture and adaptability in weak texture regions. We propose a refined optical flow estimation network integrating multiscale kernel selection and super-resolution techniques. A parameter-free Swift 3D weighted attention mechanism dynamically adjusts feature importance, while a super-resolution module enhances fine-grained feature capture. A context network with multiscale kernel selectivity expands the receptive field, improving adaptability. Experimental results demonstrate significant accuracy improvements on datasets like Sintel and FlyingChairs, showcasing the method's effectiveness in complex scenes. Code location https://github.com/yifanna/PS3D-flow.git.