93db85ed909c13838ff95ccfa94cebd9-AuthorFeedback.pdf

We also compare to Gibbs in terms of32 accuracy(Appendix, Figure 4,middle) andruntime (right). Foradatapoint where thesource only assigns frame-levellabels (say allpositive),theoutput vector would look like43 Y = [1,1,1,0], where 0 is an abstain. In case it applies a label toF1 and the overall scene, it would look like44 Y = [1,0,0,1].