How are the features of arms and legs used?
quan1e opened this issue · 5 comments
Hello authors:
Sorry to bother you,
The demo shows a great visual effect, while I have some simple puzzles about it.
It can be seen in the demo, the arms are divided into upper and lower arms, and sometimes their boxes are rotated. So, did u use the features in the tiny rotated arm and leg boxes to complete downstream tasks?
And how can the features in rotated boxes be used?
Thanks.
The rotated arm and leg boxes are only used for the presentation of human pose, and are not used to extract features. Actually we generate ten part boxes(not rotated) following previous work Pairwise Body-Part Attention for Recognizing
Human-Object Interactions, and extract the RoI-Pooling features of the boxes for the downstream tasks.
For the details, please refer to the implementation in this repository.
I read the file part_box_generation.py. So, the part boxes are generated with the rules in it. Then train the part model with pre marked labels and the features in boxes.
BTW, what do f_parts, s_parts, p_parts mean in the code?
f_parts are the features of part state classification, which perform like the outputs of fc7 in VGG-16.
s_parts and p_parts are the predicted logits and sigmoid scores of part states.
Oh, I see. Thanks for your swift relpy.