Regarding the person IDs

[bmmtstb]

After I was getting errors in my one hot person ID classifier, I decided to investigate the amount of person IDs. And I'm quite confused now...

I set up a classifier using the 5474 IDs mentioned in the Person Search Dataset description. So the amount of unique values is correct, but its written nowhere, that the numbers are not in range 0..5474.

The following table shows the min and max value of the person IDs, and the amount of unique values in the directories of posetrack_person_search and posetrack_data:

Dataset              Min - Max  - Unique
Search - Train:       11 - 7529 - 5474
Search - Query:        3 - 1600 - 1313
Search - Val/Gallery:  0 - 1655 - 1656
Search total:          0 - 7529 - 7129
---
PTrack - Train:     1600 - 6878 - 4172
PTrack - Val:          0 - 1601 - 1526
PTrack total:          0 - 6878 - 5696
---
All:                   0 - 7529 - 7130

Now I have a few questions:

• Are the person IDs between PTrack and Search "the same" / interchangeable? Meaning that ID 1 of PTrack refers to the same person as ID 1 of Search?
• To not have a classifier with empty classes, I assume one should create a mapping from real pIDs to a zero-indexed version without holes? Is there an existing mapping?
• What should the number of classes be for the classifier? I assume 7129, but then the classifier could only train ~75% of the IDs on the train set.
• I guess there is an additional testset somewhere, are there even more IDs in there, that we don't know of?

If wanted, I can publish the code I used to create those numbers, even though it is fairly straight forward using set()