Description: 

       The above is a mosaic picture from a 300-frame video clip of  tennis playing. I am sure  the picture will look nicer with suitable post-processing. This is just to illustrate how good or bad a SSD (sum of squared distance) [1] based open-loop video mosaicking algorithm can perform. The video is named "stefan" and widely used in video compression community. 

Note our method is not designed to handle dynamic scenes, but it turns out to work ok by adaptively pixel selection [2,3].

        Based on requests, I list the two core  C/C++ files (SSDUpdate.cpp & WarpingUpdate.cpp) of SSD mosaicking algorithm. Hopefuly they can help others to understand the algorithm better. Before your own C/C++ implementation, I would suggest you to build your mosaicking prototype using Matlab first. It will be extremely easy for implementation and visualization of the result.

Reference:

  1. Gregory D. Hager, P. Belhumeur, Efficient Region Tracking With Parametric Models of Geometry and Illumination, IEEE Transactions on Pattern Analysis and Machine Intelligence, 20(10), pp.~1125-1139, 1998.
  2. Xiangtian Dai, Le Lu, Gregory Hager,  “Real-time Video Mosaicing with Adaptive Parameterized Warping”, Demo Program, CVPR’2003: IEEE Computer Society Conference on Computer Vision and Pattern Recognition, May 2003, Midson, USA.
  3. Le Lu,  “Real-time Video Mosaicing with Adaptive Parameterized Warping”, Computational Interaction and Robotics Lab, Computer Science Department, Johns Hopkins University.

Comments: lelu@cs.jhu.edu