Estimating Camera-Motion Flow from Gyroscope and Optical Flow

Camera-motion flow captures image displacement caused by camera rotation and translation while excluding independent object motion. We estimate this flow by refining gyroscope-derived rotation flow using optical flow from a pretrained estimator. Two learned scores guide reconstruction from the residual between these two flows: motion compatibility identifies useful motion evidence, and scene proximity indicates nearness to the camera. Distant, compatible residuals guide global homography correction; nearby, compatible residuals guide neural parallax recovery. Our method achieves state-of-the-art accuracy on GHOF and GHOF-Cam, reducing point matching error and endpoint error by 15.4% and 28.3%, respectively, relative to the best baseline on each benchmark. Zero-shot transfer of the camera-motion modules to MPI-Sintel improves estimation over both input flows, including on dynamic pixels. Applications to video stabilization and robotic navigation further demonstrate improved stabilization and camera-pose-based reconstruction over baseline methods.