Friday January 20 , 10h00 - 11h00 , room F107 , Seminar

Georgios Evangelidis (Signal Processing and Communications Laboratory, University of Patras, Greece)

Spatio-temporal alignment of videos recorded by moving cameras

Video alignment aims at registering two (or more) video sequences in both spatial and temporal dimensions. The problem becomes quite challenging when the cameras are moving and the videos are acquired at different times.

I will present a method that permits online video synchronization and deals with highly dynamic temporal shifts. The novelty of this method lies in the adaptation of an efficient retrieval framework to a video synchronization problem. The retrieval process builds on short descriptors while a multi-querying (multiscale) scheme leads to accurate synchronization results. In addition, an image alignment scheme is extended to temporal dimension that simultaneously allows for spatial registration and synchronization refinement with subframe accuracy.

Next, I will present an alternative method that exploits similarities between spatio-temporal slices. Slice matching provides temporal correspondences by analogy to image matching that provides spatial correspondences. A Gaussian Mixture Model in conjunction with a Bayesian network build on putative temporal matches and provide synchronized frames that can be further registered in spatial domain.