Sequential image stitching for mobile panoramas

This paper presents a sequential image stitching approach for creating high-quality panoramic images on mobile devices. In this approach, each source image in the image sequence is stitched onto the panoramic image sequentially using two operations: optimal seam finding and transition smoothing. In the seam finding process, graph cut optimization finds an optimal seam and creates labeling in the overlapping area between the current panoramic image and the current source image. The current panoramic image can be updated by merging the current source image using the labeling information. If there are visible stitching artifacts in the seam, a transition smoothing operation is performed to hide the seam and remove the stitching artifacts. In the transition smoothing process, a gradient vector field is created from the gradients of corresponding pixels in the current labeled source image to construct a Poisson equation. A composite image can be recovered from the gradient vector field by solving the Poisson equation with boundary conditions. The current panoramic image is updated by merging the composite image. The approach presents several advantages. The use of graph cut optimization guarantees finding optimal seams and avoids blurring and ghosting problems caused by objects moving between capture of input images or by spatial alignment errors. The gradient domain transition smoothing process reduces color differences and further improves image quality. The sequential panorama stitching procedure enables us to produce high resolution panoramic images with limited memory resources. The approach is implemented and it produces high quality panoramic images on mobile devices. It shows good performance for both indoor and outdoor scenes.