search

verlab / accelerated_features

Implementation of XFeat (CVPR 2024). Do you need robust and fast local feature extraction? You are in the right place!
https://www.verlab.dcc.ufmg.br/descriptors/xfeat_cvpr24
Apache License 2.0
984 stars 104 forks source link

When there is a significant difference in the perspective of the images, the image matching effect is not good #55

Open canquanshen opened 2 months ago

canquanshen commented 2 months ago

matches7 The effect is not good when there is a large difference in viewing angles The following image shows the effect when the viewing angles are similar matches10

guipotje commented 2 months ago

Hi @canquanshen, thanks for the interest in XFeat! Its indeed a difficult image pair, I would recommend a more expensive but more robust approach such as RoMA, however, if compute budget is limited, there are some solutions that may help, such as multi-scale inference (it appears that there are a significant scale difference between the two images).

canquanshen commented 2 months ago

Thanks for you response. I want to know why XFeat not support significant scale difference between the two images?

zw-92 commented 2 months ago

Thanks for you response. I want to know why XFeat not support significant scale difference between the two images?

This is because the network is relatively simple, and feature matching relies heavily on the similarity matrix between descriptors. The core of the paper is that the network is simple and fast. If you need better results, you can use it ...