Closed Moreland-cas closed 11 months ago
The gain of accuracy mostly comes from fact that LightGlue's training is much more stable than SuperGlue's thanks to improvements in the architecture: softmax+matchability head replacing Sinkhorn+dustbin, relative positional encoding, bi-directional cross attention. Point pruning has negligible effects on the accuracy.
The gain of accuracy mostly comes from fact that LightGlue's training is much more stable than SuperGlue's thanks to improvements in the architecture: softmax+matchability head replacing Sinkhorn+dustbin, relative positional encoding, bi-directional cross attention. Point pruning has negligible effects on the accuracy.
Thanks for your detailed answer. It is unbelievable that there is so much room for improvement in the optimization matching algorithm without changing the position of the feature points. Do you think the consistency issue of feature points is still an important issue for keypoints-based matching methods?
Do you think the consistency issue of feature points is still an important issue for keypoints-based matching methods?
Yes, even DISK is not that good, there is a lot more work to do in this direction.
Do you think the consistency issue of feature points is still an important issue for keypoints-based matching methods?
Yes, even DISK is not that good, there is a lot more work to do in this direction.
Got it :)
Great work! Do you have an idea why the accuracy is becoming better? Is it because you exclude the irrelevant points so that the remained ones won't be affected? Any other potential reasons?