search

vcg-uvic / lf-net-release

Code Release for LF-Net: Learning Local Features from Images
Other
314 stars 66 forks source link

description match #34

Closed szgy66 closed 2 years ago

szgy66 commented 2 years ago

hello, I had run your code successfully, even use my datasets. I got much keypoints and corresponding descriptions. Just as in your paper, I want to establish the corresponding relationship between the two graphs, but you emphasize that ratio test cannot be used, so I want to ask, is there any way to match the descriptor?Or how do you filter good matching feature points.

Thanks

etrulls commented 2 years ago

You can use the ratio test. Just try different values (on a validation set) instead of picking what was recommended by SIFT. You also probably want to use bidirectional matching.

szgy66 commented 2 years ago

You can use the ratio test. Just try different values (on a validation set) instead of picking what was recommended by SIFT. You also probably want to use bidirectional matching.

I had tried. But I found that result less than sift,can't get the result as shown in paper

etrulls commented 2 years ago

What dataset are you using? What models are you using?

szgy66 commented 2 years ago

What dataset are you using? What models are you using?

Random images taken from a camera. Is there any requirement for this image. And I used the pre_trained model.

etrulls commented 2 years ago

Ok, I was confused because I thought you meant you couldn't reproduce the results in the paper. You meant it doesn't work well. I can't guess without looking at the images/results. Does it fail catastrophically, or does it just not look very good? How are you comparing it with SIFT, qualitatively? Are you using the rotation-invariant models if your data contains rotations, and the rotation-agnostic models otherwise?

szgy66 commented 2 years ago

Ok, I was confused because I thought you meant you couldn't reproduce the results in the paper. You meant it doesn't work well. I can't guess without looking at the images/results. Does it fail catastrophically, or does it just not look very good? How are you comparing it with SIFT, qualitatively? Are you using the rotation-invariant models if your data contains rotations, and the rotation-agnostic models otherwise?

It is effective, but compared with SIFT, its performance is not so good, for complex scenes, it and SIFT are both failed. I just simply observe the result of feature point matching, as shown in Figure 3 in the paper

etrulls commented 2 years ago

I'm afraid LF-Net is a bit dated, and it probably wasn't trained on your kind of data. I'd suggest you try something like SuperPoint, R2D2, or DISK, depending on the application.

szgy66 commented 2 years ago

I'm afraid LF-Net is a bit dated, and it probably wasn't trained on your kind of data. I'd suggest you try something like SuperPoint, R2D2, or DISK, depending on the application.

ok, thanks a lot

ericzzj1989 commented 2 years ago

@szgy66 Hello, have you worked out the custom datasets for training?