Some confusion about this work about lightglue HOT 2 CLOSED

Hi @Master-cai, thank you for opening this issue!

1. It certainly helps a bit in out-of-distribution benchmarks (like HPatches), but less on MegaDepth1500.
2. The network could learn to focus more on the confidence prediction than on the matching accuracy, which, in the end, is the major thing we are interested in. This could be solved by sophisticated loss balancing, but we decided to avoid this tedious tuning by not propagating gradients, which we found to be sufficiently accurate as-is.
3. We trained the confidence classifier on both tasks, but I expect that just training it on MegaDepth is sufficient. Note that training this classifier is self-supervised, so it can be fine-tuned on any set of image pairs without ground truth labels.
4. Indeed, but this is just a difference in notation. We store the non-matchability in the extra row/column. While it is not required during inference, it is required during training (Eq. 11).
5. This is indeed not required. We will update the code, thank you for this hint!
6. The problem is that bidirectional flash attention does not exist, but running flash attention twice (with shared Q/K) is faster than running bidirectional attention. We tried to implement bidirectional flash attention with Triton but did not succeed in achieving significant speedups until now. However, we do believe that there is a big opportunity for speedups there!
7. If I remember correctly this is slightly faster (because of less strided memory accesses in softmax). I agree that they are mathematically equivalent.

from lightglue.

@Phil26AT Thank you for your quick and patient response !

from lightglue.