When I evaluate my reproduced LightGlue model on the Aachen dataset, I find that this checkpoint achieves performance close to the paper’s reported results on other datasets. However, when running evaluation on Aachen using hloc, the performance drops significantly, especially on the night subset.
I suspect that the format of the saved checkpoints might not be compatible. I remember seeing someone share a script related to this issue in a GitHub discussion or issue, but I can’t find it anymore and haven’t been able to resolve the problem.
Has anyone encountered a similar situation? I would really appreciate any advice.
When I evaluate my reproduced LightGlue model on the Aachen dataset, I find that this checkpoint achieves performance close to the paper’s reported results on other datasets. However, when running evaluation on Aachen using hloc, the performance drops significantly, especially on the night subset.
I suspect that the format of the saved checkpoints might not be compatible. I remember seeing someone share a script related to this issue in a GitHub discussion or issue, but I can’t find it anymore and haven’t been able to resolve the problem.
Has anyone encountered a similar situation? I would really appreciate any advice.