The low performance of SymmCompletion in your paper

[HongyuYann]

Dear Authors,

Thank you for your contributions to the point cloud completion community and for your comparison with SymmCompletion. However, I have several questions regarding your comparison:

1. There is a significant difference in the performance of SymmCompletion reported in your paper compared to the original. While I understand that variations in training environments can lead to discrepancies, the large difference raises concerns about its authenticity. Notably, we have received several reproduced results, such as 6.30 on the PCN dataset, from other researchers, which are quite different from your reported results.
2. In the comparison on the PCN dataset, the result (which is better than yours) from SymmCompletion differs from its original. But the results (which performed worse than yours) from other methods are cited directly, which raises concerns about fairness and intentional modification. To ensure a fair comparison, please share the details to obtain the reproduced result of 6.47. We have provided checkpoints of each benchmark on GitHub for comparison, which you can use directly. If you choose to reproduce our method, you should inquire whether there is a bug when your result is significantly less than the original ones.
3. The lack of comparison with SymmCompletion on the ShapeNet55 dataset is incomprehensible.

In summary, we would like to know the reasons and details behind the poor performance of SymmCompletion, which is important for us to improve our method. Also, it is crucial for the researchers who are following SymmCompletion.

[Wanghc233]

On the PCN dataset, the results I obtained from my own training are a CD-AVG of 6.320 and an F1 score of 0.853. There is indeed some discrepancy compared with symm, but it’s not as large as the difference reported in the paper. I’m also curious why the results of the symm algorithm in Simba differ so much.