Clarify computation of quality indices for KPConv on the SUM dataset

Ascertain the exact procedure by which the quality indices (overall accuracy and mean F1) were computed for KPConv results on the SUM benchmark when classifying points sampled from the triangulated mesh, to ensure interpretability and fair comparison with triangle-level methods.

Background

For benchmarking on the SUM dataset, the authors compare their mesh-based method to KPConv, a point-cloud method applied to points sampled from the mesh. The SUM benchmark uses area-weighted metrics for triangle-level evaluation, but it is not specified how KPConv’s point-level predictions were converted to the reported quality indices.

The paper explicitly notes uncertainty about how these quality indices were determined for KPConv and therefore uses the figures reported by the SUM benchmark paper, signaling an unresolved methodological detail that affects comparability.