junshutang
Some doubts about the results on shapenet55
I found that the results in Table1, the performance of pointr is worse than the results of PCN and GRNet, which is inconsistent with the previous papers. May I ask the author, are these results tested at the same difficulty in the PoinTr paper?
It's strange. I don't know why the CD for PoinTr on ShapeNet-55 is worse than those in previous paper. We provide the pretrained ckpt for ShapeNet-55 in official repo. Did the author re-train PoinTr by themselves?
Sorry for the late reply. We didn't re-train PoinTr. We evaluate its pretrained model on all test set on ShapeNet-55 (no partition) and test different metrics: CD-l1/CD-l2/EMD mentioned in our paper. The evaluation code is similar to VRCNet and GRNet. We show the detailed results of these metrics in the supplementary material.
