search

cvg / pixloc

Back to the Feature: Learning Robust Camera Localization from Pixels to Pose (CVPR 2021)
Apache License 2.0
735 stars 92 forks source link

Can the feature points in pixloc be used in mapping？ #14

Open HawkYJH opened 2 years ago

HawkYJH commented 2 years ago

Hello, I am a graduate student in the field of slam, and I am working on my graduation project. First of all, thank you for making such a great work open source. Pixloc is designed for relocation. Can the feature points be used in mapping? Have you evaluated it on Hpatches?Compared with deep learning feature points such as SuperPoint, R2D2, Disk, etc., what is the mapping error? Another question is that in your paper, the sfm model of superpoint+superGlue is used to generate the point cloud. If use the feature points in pixloc or take a step back, the sparse point cloud generated by the orb feature points in orbslam is completely infeasible, or will it only reduce the limited accuracy?

sarlinpe commented 2 years ago
  1. PixLoc does not detect feature points, it merely estimates a the pose of a new image given a 3D model of the scene. It is thus more similar to a pose solver than a keypoint detector, descriptor, or matcher, although any comparison is limited. PixLoc thus cannot be used for mapping. To use featuremetric errors for mapping, please check out our recent ICCV paper: https://github.com/cvg/pixel-perfect-sfm (the code will be released very soon).
  2. PixLoc can use any 3D model and is not limited to SfM. For the localization on 7Scenes, we use sparse points sampled from a 3D mesh built with RDGB SLAM (more details here). Using a sparse point cloud built with ORB-SLAM is possible, although this might indeed yield a lower localization accuracy depending on the use cases.
HawkYJH commented 2 years ago

Thank you for your reply, and thank you for making such an excellent work open source！