Rotation model based quasi-dense matching propagation for uncalibrated fisheye images

One main difference for various quasi-dense matching approaches is the different assumptions about local transformations between corresponding patches of two images. The translation is assumed as local transformation for small baseline based approaches, which can not truly reflect the distortion between fisheye patches, and the affine transformation model is usually used for highly distortion images, e.g., wide baseline images, which found to be not as stable and reliable as expected. In this paper we propose a quasi-dense matching algorithm in which a novel local distortion model named rotation transformation is assumed for fisheye images. Initial local rotation transformation for seed match is obtained by phase correlation technique while the estimate of the local transformation for new candidate matches is conducted based on the current seed match. Experiment results show that our algorithm can provide satisfactory results with characteristics of simplicity, stability and reliability.