json87
Increase matches
Hello,
Is there a way to increase matches in some way? I have around 100 images and they only produce a sparse point map of around 15.000 points. I would expect a higher amount given the large image capture of 360. What settings can be tuned?
I tried to increase max matching points, extracted points, lower min num inliers, max distance etc. but it does not help. Is there any implicit limitations? Thank you. :)
There is no limitation on the number of extracted feature points and matched correspondences. We use the default parameters in ColMap. To increase matches, you should consider decrease the distortion of ERP spherical images, such as cubic map reprojection.
There is no limitation on the number of extracted feature points and matched correspondences. We use the default parameters in ColMap. To increase matches, you should consider decrease the distortion of ERP spherical images, such as cubic map reprojection.
Thanks for your prompt response. I wish I could do that but reprojecting into cubemap creates an enourmous amount of images and converging them with colmap is very time consuming when you start to approach 1000+ images.
distortion of ERP spherical images
Have you possibly checked if this would work with SphereSFM?
https://openaccess.thecvf.com/content/CVPR2023W/IMW/papers/Gava_SphereGlue_Learning_Keypoint_Matching_on_High_Resolution_Spherical_Images_CVPRW_2023_paper.pdf
Yes, you can try it. These works consider image distortions when calculating feature descriptors.
Yes, you can try it. These works consider image distortions when calculating feature descriptors.
Do you think this is something you can implement in your repo?
https://github.com/EnriqueSolarte/robust_360_8PA/tree/master
Seems to provide 20% better camera estimation due to a different normalization technique and increasing parallax. I have not programmed in C++ and my knowledge is very low for SFM matters. Thanks man for this great repo.
Currently trying out something new where I perform "dense" feature extraction and matching and then sparse SFM reconstruction. 40 ERP images have generated about one million points instead of the typical 10000 points I would get with the same images.
But seems that it some cases, the camera poses become degenerative and things start to reconstruct into each other... :(