Non-Ghosting Stitching for Street-View Image

Parallax, relative camera motion, and illumination changes can cause artifacts in stitched images. For example, we might see ghosting effect on the pedestrian who is located at the overlapping region in a stitched image. Existing techniques struggle at dealing with the artifacts and perspective distortion. We propose an innovative image stitching algorithm which combines portrait segmentation and image inpainting. Images from different cameras are first segmented into the pedestrian plane and the scene plane by portrait segmentation. Then, images in different planes are stitched. Finally, the pedestrian plane and the scene plane are fused by aligning and inpainting. Experiments show that our method can remove ghosting effect in stitched street-view images more effectively than existing meth-ods.