Seite - 149 - in Joint Austrian Computer Vision and Robotics Workshop 2020

input SL target N2N input SL target N2N input SL target N2N Figure3.Thefirst row depicts crops fromthecorrupted framexij alongwith thecorrespondingmanually edited target y¯ i j. Thesecondand third row showthe results obtainedusing the static modelNθS, whereas, the resultsof the dynamicmodel aredepicted in the last three rows. Thecolumnsalternatebetweensupervised learning (SL)andN2Nresults andon the rightweshowwhich loss functionwas usedduring training. in which the reader was presented three versions of the samescenesidebyside: (i)Theoriginal frames, the output of the models trained using (ii) SL and (iii)N2N (‖·‖ , = 0.1). Table2 presents the results obtained from 24 people who were each shown 8 video sequences. It shows that the model trained with N2N is best at removing the defects, at the cost of over smoothing the images. Still, it was the overall preferred method,with53.65%ofall samplesbeing deemed “OverallBest”by theparticipants. 5.Conclusion In thiswork we explored the possibilities ofusing N2N learning for video restoration. We trained static anddynamicalmodelsbyconsideringadjacentframes using supervised learning and N2N, relying on robust motion estimation. Using this paradigm we demon- strated thatvideo restorationcanbe learnedbyonly looking at corrupted frames at performance levels exceeding those of supervised learning. This opens 149