Reconstructing Sharply Folding Surfaces: A Convex Formulation

Mathieu Salzmann Pascal Fua
EECS & ICSI EPFL - CVLab
UC Berkeley 1015 Lausanne, Switzerland



The following videos are encoded in DivX.

For Windows users: DivX Player.
For Apple users: QuickTime Player.
For linux/unix users: mplayer.

In each video, we show, on the top left, the mesh recovered using the motion model reprojected on the original images, on the bottom left, the same mesh seen from a different viewpoint, and, on the top right, the mesh recovered without using the motion model.
Since the motion model relates 3-frame blocks but does not enforce temporal consistency over consecutive blocks, the results are essentially correct but seem jittery when shown as video sequences. This is easily remedied by post-processing the sequence of recovered shapes to enforce temporal smoothness over larger sub-sequences, which can also be formulated as a convex problem. We applied this on 11-frame blocks, and show the results in the bottom right corner of the video. Note that this final smoothing step is not mentioned in the paper, since our approach focuses on recovery in individual frames or in very short 3-frame sequences.