| Visual stitching under challenging conditions (e.g., large parallax, weakly-textured content, and freely-moving hand-held camera inputs) poses great challenges to existing stitching solutions. To deal with these problems, we propose several state-of-the-art solutions for different situations. Firstly, we propose a seam-guided local alignment method (SEAGULL) for large parallax image stitching. Then, we introduce a mesh-based photometric alignment (MPA) method for weakly-textured image alignment. For video stitching task, we develop the first system to stitch videos from freely-moving hand-held cameras. It combines dense 3D reconstruction and a line-preserving warping method to generate spatially artifact-free and temporally stabilized stitching results. Lastly, we present a robust trajectory-based background identification method, which can be used in video stitching to robustly identify the static background features even with large foreground objects. The experiment results show that our methods can handle a variety of challenging images and videos, and outperform many state-of-the-art stitching and motion segmentation methods. |
