This repository contains files used to create my winning entry to the Celtra 2017 AI Challenge. The challenge was to predict the viewer orientation in a 360 degree video based on previous viewer orientation data. Using these predictions the video was to be broken up into non-overlapping segments with different video qualities that could be changed on the fly based on the predicted viewer orientation.
The challenge allowed for 10 submissions to the global leaderboard in total, so implementing good local cross-validation was crucial.
My solution used (linear) resampling, windowing, and seam-carving (dynamic programming) to produce the final predictions. Cross validation was successfully used to tune the parameters for submissions.
Brief description of files in the repository:
challenge_description_eng.pdf- English description of the challengedata/- contains the input train and predict datasets along with sample submissionsrc/- the source code of the solutionmain.lisp- the main entry point of the solutionfinal_*- the final leaderboard, report, and presentation (in Slovenian)
To reproduce the results, install sbcl and quicklisp, then run:
$ sbcl --eval '(progn (push (sb-posix:getcwd) ql:*local-project-directories*) (ql:register-local-projects) (ql:quickload :celtra))' --load main.lisp