Questions regarding methodologies used in "Revisiting Skeleton-based Action Recognition" paper

[liamwebsterxyz]

Hey, great paper! I just have a couple of questions that I have not been able to find answers to.

"Ensembling the results from joint-PoseConv3D and limbPoseConv3D (namely PoseConv3D (J + L)) can lead to
noticeable and consistent performance improvement."
- What method of ensembling did you use?
- Did you experiment with populating the 3D heat map volumnes with both the joint and limb information? That is a 3D J+L heat map volumn?

In the context of RGBPoseConv3D which skeletal information modality did you use e.g. joint or limb? Did you experiment with fusing (J+L) skeletal model with RGB model? What mechanism of late fusion did you use e.g. concatenation, averaging, linear layer?