Systems and methods for stitching videos are disclosed. Image-based registration between frames from a first video source and frames from a second video source is performed at a first rate. Calibration-based registration between frames from the first video source and frames from the second video source are performed at a second rate higher than the first rate. Then, for a first frame from the first video source for which calibration-based registration data and image-based registration data have been generated, a stitching transform that maps the first frame to a counterpart frame from the second video source based on image-based registration data is generated. A delta transform from the image-based registration data and the calibration-based registration data at the first frame is also derived. For a subsequent frame from the first video source for which calibration-based registration data have been generated, but no image-based registration data have been generated, a stitching transform that maps the subsequent frame to a counterpart frame from the second video source based on the calibration-based registration data and the delta transform is generated. Frames from the first video source and frames from the second video source are stitched according to their respective generated stitching transforms.