Pixel-Wise Warping for Deep Image Stitching

Existing image stitching approaches based on global or local homography estimation are not free from the parallax problem and suffer undesired artifacts. In this paper, instead of relying homography-based warp, we propose a novel deep framework exploiting pixel-wise warp field to handle large-parallax problem. The proposed consists Pixel-wise Warping Module (PWM) Stitched Image Generating (SIGMo). For PWM, obtain in similar manner as estimating an optical flow (OF). scenario, input images usually include non-overlap (NOV) regions which cannot be directly estimated, unlike overlap (OV) regions. To help PWM predict reasonable NOV region, impose two geometrical constraints: epipolar loss line-preservation loss. With obtained field, relocate pixels target using forward warping. Finally, SIGMo is trained by multi-branch training generate stitched reference warped image. evaluating framework, build publish dataset including pairs with corresponding ground truth result images. We show that results quantitatively qualitatively superior those conventional methods.