search

open-mmlab / mmskeleton

A OpenMMLAB toolbox for human pose estimation, skeleton-based action recognition, and action synthesis.
Apache License 2.0
2.92k stars 1.03k forks source link

pose_matching and temporal kernel #282

Closed imduygusarikaya closed 4 years ago

imduygusarikaya commented 4 years ago

If pose_matching is set to False, would this make only the spatial graph convolutions work on a frame basis with no temporal connections? In this case, are features from each frame just stacked together? How would the temporal information work? Or am I missing some subtle information here? It is awkward that I get better performance when it is set to false, considering that temporal information is important in action classification.

Also, is the kernel size set at 9, related to the number of frames (temporal dimension)?

yjxiong commented 4 years ago

No. If it is set to FALSE we do not tracking and assumes we always assume the skeleton data of the persons in each frame are sorted by their IDs when forming the sequences.