interacting hands criteria

[ccchen091]

Hello, I found a problem when training your model, the types of hands detected in your model are all single hand, but when I open the dataset to observe, I find that there are many interactive hands, not just one-handed pictures, so I would like to ask you what is the problem? “”“”“”“”“”“” Get bbox and root depth from groundtruth annotation Number of annotations in single hand sequences: 61153 Number of annotations in interacting hand sequences: 0 Number of skipped annotations: 387 “”“”“”“”“”“”“

[tkhkaeio]

In this codebase, we didn't assign the interacting label but rather created hand crops for each right and left hand. See this line.

You can modify this data loader according to your criteria for interacting hands. For instance,

• compute the distance between two hands in a frame
• if the dist is less than a threshold, the interacting label is assigned.
• then create crops covering the two-hand region.

I hope this can work for your case.