num_correspondences scale

crockwell / far

[CVPR 2024 - Highlight] FAR: Flexible, Accurate and Robust 6DoF Relative Camera Pose Estimation

https://crockwell.github.io/far/

101 stars 6 forks source link

num_correspondences scale #3

Closed wuqun-tju closed 5 months ago

wuqun-tju commented 5 months ago

hello, I have one question , in th code , num_correspondences is divided by 500, 10 or 100， I want to know how to choose this constant scale?

crockwell commented 5 months ago

Good question! So there are 3 values for number of correspondences. (1) The number of inliers (2) the number of inliers at a tighter threshold (3) the number of inliers at an even tighter threshold.

We divide all 3 by 500 to make them of reasonable scale for the network (here)
(2) and (3) are inliers at thresholds of 1/10 and 1/100 the normal threshold, and are defined here. Their values aren't actually divided by 10 or 100, but rather their threshold is just different.

Hope that helps! Chris

wuqun-tju commented 5 months ago

Thank you for your reply ! I still have confusions as follows: (1) the number 500 can be replace with another number？ I want to know how to chosse the number value (2) 10 and 100 is used here

crockwell commented 5 months ago

(1) Sure, the number is arbitrary, picked to try to match ranges of other network outputs. I found this number worked well. (2) The reason I did this was a bit ad-hoc but I wanted the network to be trained with 3 input number of correspondences, even though only one number is available when using OpenCV RANSAC, which we use in early training of FAR (i.e. T1). 3 are available when using the full network, as the Kornia function is custom and can return three values. To get the network prepared for the range of values expected, I divided by 10 and 100 if only the OpenCV number of correspondences were available.