Non animal key point models - I'm using key points to find the pose of pallets...

[pgaston]

I'm interested in finding the key points of a pallet. With this information and simple geometry, I can somewhat accurately find the pose of the pallet. Great. (And yes, there are other ways we're looking at, e.g., direct to pose...)

I've shown success using a few models. Initially Detectron2, and now, having switched to mmPose I've used in the first week CID and RTMPose.

Questions:

• Is there anything inherently wrong in using key points for non-animal objects? It seems to work w/ pallets.
• I do see roughly 5% errors where key points are strangely assigned to the wrong pallet. See image below. CID claims to be good w/ multi-person - I had hoped this wouldn't happen here. But it also happens w/ Detectron2. (Not w/ RTMPose - but that seems to be one object at a time...)
• Bottom-up vs. Top-down. Any general rules of thumb regarding accuracy and speed? i.e., should I prefer one over the other?

Thanks - I'm off playing w/ various models! p

First a bad one, then a good one (the high majority)