search

open-mmlab / mmskeleton

A OpenMMLAB toolbox for human pose estimation, skeleton-based action recognition, and action synthesis.
Apache License 2.0
2.94k stars 1.04k forks source link

spatial partition strategy in graph.py #245

Open hsuankai opened 5 years ago

hsuankai commented 5 years ago

For spatial partition strategy, the centripetal group is defined by rj < ri in paper, which means the distance between the center and neighboring node j is shorter than the center and node i. However, in graph.py, the close group is defined by self.hop_dis[j, self.center] > self.hop_dis[i, self.center]. To my opinion, it should be self.hop_dis[j, self.center] < self.hop_dis[i, self.center], because I think that a_close represents those neighboring node j which are closer to the center node. Is it correct?

thancaocuong commented 5 years ago

@hsuankai hi, can you explain me why choose self.center = 21 -1 in ntu-rgb_d dataset? in their paper, they said that "the average coordinate of all joints in the skeleton at a frame is treated as its gravity center". But I dont how why they choose 21 - 1 . thank you