טכניון מכון טכנולוגי לישראל
הטכניון מכון טכנולוגי לישראל - בית הספר ללימודי מוסמכים  
M.Sc Thesis
M.Sc StudentElbaz Romi
SubjectAutomatic Looping Video Synthesis
DepartmentDepartment of Electrical Engineering
Supervisor Professor Lihi Zelnik-Manor
Full Thesis textFull thesis text - English Version


Abstract

In recent years the problem of converting an arbitrary video into a looping content has become a challenging research problem. The exploration of this field is motivated by the notion that such an intermediate medium between video and static image can better capture the "moment" and convey subtle motion in a scene to make it come alive. In addition, it has been claimed that the contrasting juxtaposition of looping elements against still backgrounds helps grab the viewer’s attention.

In this research previous methods for video looping are reviewed. These solutions can be categorized into two main approaches. The first approach offers the user a creative tool to selectively freeze, play, and loop video regions to achieve compelling effects. These short dynamic videos are named cliplets or cinmagraphs, and have recently become of great popularity. The second approach focuses on fully automating the process of forming looping content from short video.

Fully automated approaches were first offered by Schödl et al[7] who attempted to render videos that appear to play continuously and indefinitely, by selecting the optimal frame from the input video at each time point. Next, Agarwala et al. [1] offered a different approach that is semi-automatic. Rather than considering the transition between entire frames in the input video, the optimization was done separately for different looping regions in the video to better cope with scenes that contain a variety of dynamically looping regions. The partition to regions however, is user- defined, hence this method is only semi-automatic. Finally, Liao et al [6] reduced the scope of the optimization process to optimize the temporal mapping of each and every pixel coordinate in the video.

In this research a fully automated coarse-to-fine approach is proposed to try and reduce the computational cost of the previously proposed pixel based approaches.