Selfie Stitch: Dual Homography Based Image Stitching for Wide-Angle Selfie

Sourabh Yadav, Pradeep Choudhary, Srishti Goel, Sankaranarayanan Parameswaran, Pankaj Bajpai, Jaehyun Kim

Abstract: This paper proposes an image stitching method for increasing the Field-of-View (FoV) of widely used face photographs called ‘selfies’. Unlike traditional methods that deal with the far-field object and single dominant plane homography, we focus on the near-field aspect of the face and two dominant planes. In this context, significant depth difference between the face and a background plane poses significant parallax challenges. To mitigate this, we automatically calculate robust homographies for each plane by segmenting images into the foreground and background regions. To address the issue of holes caused by dual homography based stitching, we estimate a per-pixel homography by weighting the foreground and background homographies using distance transform from segmentation boundary. The experimental results show that the proposed method generates aesthetically stitched face photographs under the natural environment with reasonable post-processing time.