Pose_cfg Parameters

Hi, my team and I are working on a project using maDLC, and in an effort to make the best use of the software we are trying to understand the details of how it works. To this end, I’ve been looking into the parameters in pose_cfg.yaml to try to understand what they do, but I’ve been having trouble with some of them.

Some of them I haven’t really found anything about:
pairwise_huber_loss
pairwise_loss_weight
pairwise_predict
metadataset
weigh_only_present_joints

For these ones I was directed to the paper DeepCut: Joint Subset Partition and Labeling for Multi Person Pose Estimation, but I’m still not really sure exactly what they do:
location_refinement
locref_huber_loss
locref_loss_weight
locref_stdev

I’m also a bit unclear about what pafwidth and partaffinityfield_predict do. And what exactly is a part affinity field?

Finally a couple clarification questions: my interpretation of pose_dist_thresh is that any joint marked by the program within this distance of the human defined ground truth is considered a successful detection. Is this correct?
And for rotation, it says this is only used if the training set is symmetrical around a vertical axis. Does this mean that the animals must have matching left and right body parts, or is there more to it? Does it matter if the animal is on its side in some frames?
And for the intermediate supervision parameters, does that only add one loss layer, or multiple? And what does intermediate_supervision_layer determine?

Any help would be very much appreciated! Thank you!