search

cvlab-epfl / segmentation-driven-pose

Segmentation-driven 6D Object Pose Estimation. CVPR 2019.
GNU General Public License v3.0
184 stars 37 forks source link

Coordinate system convention of the ground truth poses #5

Closed SelvamArul closed 5 years ago

SelvamArul commented 5 years ago

Hi, One more question related to original LineMOD. The ground truth poses in Occluded-LineMOD dataset follows the OpenGL Coordinate system convention (camera viewing direction is the negative Z-axis) as mention in the section 2.2 of this document. But I am not sure of the coordinate system convention used in the original LineMOD dataset. For example one of the transformation file like looks like -13.0792 -7.83575 104.177 The values are in centimetres. so, that's ~1 meters along the positive z-axis. This make me think that the poses are annotated with OpenCV coordinate system convention. But my rendering pipeline is not working correctly if I follow this assumption. Any idea how the ground truth poses in original LineMOD dataset is defined? Also, what this file in your repo is for?

Sorry that this issue is not about the segmentation-driven-pose repo but rather a general question.

YinlinHu commented 5 years ago

This do is a general confusing problem. As there are many versions of LINEMOD dataset, we use the normalized one at: https://files.icg.tugraz.at/f/472f5d1108/ But as you may notice, the model used in Occ-LINEMOD (https://hci.iwr.uni-heidelberg.de/vislearn/iccv2015-occlusion-challenge/) is not the same as the above. While, only a simple model transformation is enough, that is what the file you mentioned exactly do.

SelvamArul commented 5 years ago

Thanks a lot for your reply. I am using the normalized version of LINEMOD you shared and it works. BTW, where did you find the link for downloading normalized LINEMOD or any mention of it? The project page from ICG deos not mention anything about it. As you said, there are many different versions of LINEMOD and I came across many projects here in github that use LINEMOD but none mentioned this version. Is there any paper that mention this version? I am sorry if I missed something obvious. Thanks again for your reply.

YinlinHu commented 5 years ago

We just got the dataset from the member of ICG, it obviously is a cleaned and refined version, but without any documents. While, which version you use is not that important, just make them consistent is enough, I think.

SelvamArul commented 5 years ago

Ya, sure. This normalized version is much cleaner. I will stick to this version.