Within
https://shapenet.cs.stanford.edu/iccv17/partseg/part_seg_eva.m, which has been used for later analysis with PointNet++ , RelationShape for the best of my knowledge.
If there are no parts, then we get (0 + eps) / (0 + eps) = 1.
It augments the metric score, and some paper might just skip parts without any semantic labels, which makes more sense.
I am currently undergoing the re-implementation of all SOTA models on all datasets and that seem to be wrong to us:
https://github.com/nicolas-chaulet/deeppointcloud-benchmarks
Could it be possible for anyone to review and contact me at thomas.chaton.ai@gmail.com