percision in OPV2V dataset

[KuMyy]

hi,
I trained CoAlign using the OPV2V dataset with the configuration file pointpillar_coalign.yaml, but the resulting accuracy is far from what you reported. My results are as follows:
AP@50
0.953
0.947
0.932
0.911

AP@70
0.869
0.841
0.809
0.776

I downloaded your provided model and configuration files for testing, and indeed, those results match what you stated in your paper. I compared my training setup with your provided config.yaml file, and they are identical—meaning the model architecture and hyperparameters are exactly the same.

Given that, why is there still a noticeable performance gap? Did you apply any additional processing or tricks during training? I'm using an RTX 3090, so hardware differences shouldn't be the issue.