Lots of visual recognition problems have a multi-label nature. That means an entry (Datum) will be assigned to multiple labels. Currently the DataLayer and LevelDB converter assume there is a single label assigned to every entry.
Has anybody worked on that yet?
My suggestion is that we make an alternative "label array" \in {-1,0,1}^k where k is the number of possible labels and
-1: label does not exist in the entry image
0: undefined/unknown/irrelevant label for the image
1: label exist in the entry image
Lots of visual recognition problems have a multi-label nature. That means an entry (Datum) will be assigned to multiple labels. Currently the DataLayer and LevelDB converter assume there is a single label assigned to every entry.
Has anybody worked on that yet?
My suggestion is that we make an alternative "label array" \in {-1,0,1}^k where k is the number of possible labels and
-1: label does not exist in the entry image
0: undefined/unknown/irrelevant label for the image
1: label exist in the entry image