Multi-viewpoint Panoramic Stitching For Wide-baseline Images

With the prevalence of digital cameras and smart phones, sharing captured and collected images has become a popular way to entertain people and brand ourselves. Since a normal camera has very limited field of view, most applications provide a panoramic shooting mode, where the users can capture images under guidance to generate a panorama. Although panoramic stitching has been well studied for single-point view perspective model, it is still quite a challeng in the multi-viewpoint scenario.However, in the mobile applications, it is difficult to take photos in a strictly rotational way. With objects close to the camera, the baseline can be large enough to break the assumptions for most of the existed methods. Further more, in a security camera systems, cameras are scattered over the scene, it would be a great help to stitch those camera views to a single panorama. In such cases, multi-viewpoint panoramic stitching is desirable.We present a novel image stitching approach, which can produce visually plausible panoramas with images taken from different viewpoints. Unlike previous methods, our ap-proach allows wide baseline between images and non-planar scene structures. Instead of 3D reconstruction, we design a mesh-based framework to optimize the alignment and regulari-ties in a 2D sense. By minimizing a global energy function, we can construct a high-quality panorama which is locally perspective but orthogonal as a whole. Meanwhile, we improved the previous methods for seamless composition and achieved better performance on mis-aligned areas.