Temporally consistent disparity maps from uncalibrated stereo videos

We address the problem of computing a sequence of dense disparity maps from two synchronized video streams recorded by slightly displaced cameras. Generating such disparity videos is becoming increasingly important in the light of new autostereoscopic displays and novel viewpoint applications. We propose a good-quality, computationally fast and easy-to-use solution to accomplish this task. This paper describes the four major steps of our 2D to 3D video conversion procedure. (1) The user segments the video into its scenes. (2) For each scene, we rectify the uncalibrated stereo pairs so that correspondences lie on the same horizontal scanline. (3) A fast and accurate dynamic programming-based stereo matcher is then applied to compute a dense disparity map for each stereo pair. (4) We perform temporal smoothing on the computed disparity sequence to reduce the disparity flickering problem. All of these functionalities can be accessed via an easy-to-use graphical user interface, which makes our conversion procedure applicable even for technical unskilled users. We demonstrate the good quality of our results using various challenging real-world stereo streams.