Inquiry about accuracy comparison between 4-heads and 8-heads configuration with 512 keypoints

[knightXun]

Hello LightGlue team,

First, thank you for your great work on LightGlue! It's an impressive and efficient feature matching solution.

I have a question regarding the model's configuration when dealing with a moderate number of keypoints (specifically 512 keypoints). I'm trying to understand the specific impact on matching accuracy when reducing the number of attention heads from 8 to 4 in this scenario.

My specific questions are:

1. In LightGlue with 512 keypoints, is there a measurable accuracy difference between 4-head and 8-head configurations?
2. If yes, approximately how much accuracy reduction should we expect when using 4 heads instead of 8 heads?
3. Are there any known trade-offs specific to this keypoint count where the benefits of 8 heads might diminish?