What does 'position' exactly mean in the annotation?

#2
by Becomebright - opened

Hi,
Your paper mentioned that you “ask annotators to explicitly label the frame index for all referred moments in each question.” I believe the ‘position’ attribute in the validation annotations corresponds to the frame index. Could you provide a clear explanation of this? It’s confusing because different samples have varying numbers of positions.

Becomebright changed discussion title from Can you provide the labeled frame indices of referred moments? to What does 'position' exactly mean in the annotation?
LongVideoBench org

Yes, the position is the frame index. For L1 questions, as they only correspond to one moment of the video, they only have one position. However, for L2 questions, they are related to multiple moments in the video (e.g. asking about the sequences of two events, which are in in 3:05 and 6:30 respectively), so all related positions are labeled, leading to more than one position (frame index).

Hope this clarifies.

Sign up or log in to comment