Closed TWang1017 closed 2 years ago
COLMAP's pose does not have a metrical scale from pure RGB images. RealSense has a metric scale so the trajectory is in meters. To align them, you may want to use point to point registration and turn on with_scaling
(http://www.open3d.org/docs/release/python_api/open3d.pipelines.registration.TransformationEstimationPointToPoint.html). This will give you the approximate scale between two trajectories and thus metrics.
COLMAP itself doesn't guarantee the reconstruction accuracy so distortions are more likely inevitable from the pipeline, I wouldn't think calibrating the images will make a big difference.
COLMAP's pose does not have a metrical scale from pure RGB images. RealSense has a metric scale so the trajectory is in meters. To align them, you may want to use point to point registration and turn on
with_scaling
(http://www.open3d.org/docs/release/python_api/open3d.pipelines.registration.TransformationEstimationPointToPoint.html). This will give you the approximate scale between two trajectories and thus metrics.COLMAP itself doesn't guarantee the reconstruction accuracy so distortions are more likely inevitable from the pipeline, I wouldn't think calibrating the images will make a big difference.
Thanks you very much for the hints, will try it out!
Checklist
master
branch).My Question
Hi, I collected rosbag file with realsense d455 and then used the open3d reconstruction system pipeline to produce a model.
In the meantime, i found the trajectory.log file in the scene folder which i reckon is the reference for the reconstruction pipeline in open3d.
I then tried to make some comparison with image-based method, colmap and openmvs etc with the known poses from open3d (pose from trajectory.log, intrinsics from intrinsic.json). I inverted the trajectory as i need extrinsic (i.e world to camera).
The extrinsics and intrinsics have experimented in the colmap with point_triangulator to produce a sparse model for inspection before dense reconstruction and it succeeded with two issues. The general shape is right. The white is produced by open3D and the sparse model in red is produced by colmap
I have been stuck on the scale and pyramidal issue here for one week, any leads and advice are much appreciated.