Open AmokraneIlhem opened 1 year ago
@liuyuan-pal Can you please explain me more how to use colmap to annotate the ground-truth pose for evaluation.
You may need to label at least 3 points on two images from the query sequence and two images from the reference sequence Then, you can use triangulate to compute the 3D coordinates of these 3 points in the coordinate system of the query sequence and their coordinates in the coordinate system of the reference sequence. Then, you can compute the similarity transformation from these 3 points. In this case, this similarity transformation will align two sequences and you can known the labelled pose, which is regarded as gt poses.
@liuyuan-pal
Are the steps you mentioned, i.e., triangulation and similarity transformation already implemented in the code, or do I have to implement them myself?
You can implement it yourself, triangulation would be done by an opencv function while estimating similarity transformation would be an SVD decomposition.
@liuyuan-pal Can we use cloudCompare to label these points or are there other tools to do so? can you give me please the steps to follow to use the colmap .
@AmokraneIlhem Can we use cloudCompare to label these points or are there other tools to do so? can you give me please the steps to follow to use the colmap . Hello collaborators, may I ask, have you successfully tested the accuracy on your own data set?
Hello @liuyuan-pal, does the "compute_align_poses.py" script calculate the accuracy of the object's pose? If yes, can I use it for custom objects?
You have to annotate the ground-truth pose for evaluation. I did it by COLMAP retrieving the relative poses and then manually selecting points to register the reference sequence with the query sequence.