Piecewise planar super-resolution for 3D scene

For super resolution (SR) of a 3D scene from multi-view low resolution (LR) images, subpixel registration is one of the main problems due to the difference of the depth for different objects, especially for hand-held uncalibrated cameras. In this paper we proposed a method to solve this problem by estimating the depth map for the LR images and treating the registration as a planar segmentation problem. Thus we can estimate the homogeneous transforms for each plane among all the LR images. Besides that, the occlusion problem can be well handled according to the depth information. Moreover, we find an approach that combines both the multi-view reconstruction and single view reconstruction. By using the result of multi-view SR reconstruction, we can plug in the high frequency information via dictionary learning method. The experiments show that the combination of the multi-view and single view reconstruction performs better than both these two individual methods.