Rectangular-output image stitching

Image stitching aims to combine two images with overlapping fields to expand the field-of-view (FoV). However, the stitched images of existing methods are irregular, and need to be processed by rectangling methods, which is time-consuming and prone to be unnatural. In this paper, we propose the first end-to-end framework, Rectangular-output Deep Image Stitching Network (RDISNet), to directly stitch two images into a standard rectangular image while learning color consistency between image pairs and maintaining the authenticity of the content. To further preserve the structure of large objects in the stitched image, we design a dilated BN-RCU block to expand the receptive field of RDISNet for extracting enriched spatial context. Furthermore, we design a novel data synthesis pipeline and build the first rectangular-output deep image stitching dataset (RDIS-D) for jointing image stitching and rectangling. Experimental results demonstrate that RDISNet performs favorably against the state-of-the-art methods.