Aligning Time Lapses of Stars Using Contextual Information
Authors:
Holly Chu, Justin HoogenstrydMentor:
Ernie Esser, Post-Doctoral Scholar of Mathematics, University of California IrvineWe propose a method of video stabilization for a video time lapse of stars by utilizing a contextual shape descriptor to match corresponding stars on different frames. The descriptor is based from Belongie, Malik, and Puzicha [4] using log-polar bins to mask onto an image. Contextual landmark tracking is not well used in video stabilization because of it’s inefficiency and computational expense. By incorporating rotational invariance, contextual landmark tracking can possibly be proved to be worth the expense for some examples and extend applicability. While landmark tracking is not commonly used, it can be proved useful if purely local information is insufficient. Our challenges with the data were due to points disappearing and reappearing in different frames, rotating off the frame, and also that each point looked alike. In terms of our rotationally invariant descriptor, we looked into Bhonsle and Klinsman [2] for their uses of the Fast Fourier Transform. Our rotationally invariant descriptor takes an initial frame and identifies the same point in another frame whose points are completely rotated from the original. When we make that match, it will prove that our solution is effective.