On Improving Space-Time Super Resolution Using a Small Set of Video Inputs

We address the problem of video super resolution - obtaining a single high-resolution video whose spatial and temporal resolutions are higher than any of the input low-resolution videos. We employ a reconstruction based approach using MRF -MAP formalism, and use approximate optimization using graph cuts to carry out the reconstruction. We show that in our formulation it is possible to selectively apply super resolution constraints to low resolution pixels to contribute towards space or time, depending on whether they appear in the static or moving parts of the scene without any prior segmentation. As a consequence, we can achieve space-time magnification factors beyond what is suggested as possible in [19]. We also provide insights on how to keep the super resolution reconstruction problem well-conditioned.