I think (mpIoU, 84.6 in PointCNN) is same as (mean, 85.1 in Pointnet++), as in both cases the IOU is calculated for each sample (part-averaged) and then averaged over the dataset. Further, pIOU metric does not exist in PointNet++. The (mean shape IOU, 81.9 in PointNet++) is averaged across the shape categories and not individual parts.
Please correct me on this if I am missing something. I would very much appreciate some clarifications here.
Hello,
I really like your paper and thanks for sharing the code. However, I have some questions about the evaluation metric used for ShapNet Part Segmentation. Looking closely into the evaluation metric used in PointNet++(https://github.com/charlesq34/pointnet2/blob/master/part_seg/evaluate.py#L166-L182) and PointCNN(https://github.com/yangyanli/PointCNN/blob/master/evaluation/eval_shapenet_seg.py#L80-L104), I feel the numbers are not comparable.
I think (mpIoU, 84.6 in PointCNN) is same as (mean, 85.1 in Pointnet++), as in both cases the IOU is calculated for each sample (part-averaged) and then averaged over the dataset. Further, pIOU metric does not exist in PointNet++. The (mean shape IOU, 81.9 in PointNet++) is averaged across the shape categories and not individual parts.
Please correct me on this if I am missing something. I would very much appreciate some clarifications here.
Best, Ankit