fix existing labelled data (cutfiles) #15
Description
1. noisy labels
the annotated program boundaries are noisy. Therefore, using it as ground-truth may not give useful loss signal in a supervised learning setting.
the pb annotations in the cutfiles do not seem to consider commercials.
say:
program_A ->(pb#1) commercial ->(pb#2) program_B
(pb#1) and (pb#2) should be considered different and both should be annotated.
However, the existing cutfiles annotated one of (pb#1) and (pb#2) and likely in an inconsistent manner.
2. is this a valid classification task? aren't samples correlated in time?
how should I deal with this?