Loss of precision due to confidence module

[QileSuu]

Thanks for your contribution in picture matching, which brings a lot convenience for my job. However when I tested this LightGlue model
on my dataset, I focused that the same data exhibited significant variability across different batches, which was obviously non-sense. After checking the details of the structure of codes, I figured out that this error might be caused by the confidence module which seemed to use the average confidence across all batches as the evaluation criterion. I am considering that it would performs a better accuracy if this module can be somehow improved. Due to time constraints, I haven't delved deeply into the code, so my viewpoint may be flawed. Please point out any issues if you notice them. I hope this helps you.