About Feature Points Pixel Accuracy

[cho-wang001]

Thank you for your work!

I found that the coordinates of the feature points in the left image are all int (integer) values of pixel when trying to match the feature points of the two images. But the pixel values on the right are float.

I hope there is a potential solution. Thank you.

[shuow2]

I feel that's how the method is designed to be.

See "3.4. Coarse-to-Fine Module" of the paper.

mkpts0_f are simply copied from mkpts0_c, so they stay at coarse pixel positions (integers)

mkpts1_f are refined to subpixel accuracy — floats, through the fine-level attention + correlation refinement.