Confusion about PoseNet architecture about sfmlearner-pytorch HOT 2 CLOSED

short answer, because the original author did so

Long answer, we don't want the relative pose between two, but between target frame and each reference frame. We could indeed do a network that takes two frames into input the target frame and one reference, and do that for every reference frame, but that means the pose network only have temporal context of two frames. The fact that all the frames in the sample are temporally close to each other would not be leveraged.

The take away point is that computing pose of a whole sequence of frames is deemed easier than pose of simple pair.
Now you can probably argue that "2frames" pose networks are not that and some papers tend to agree with this http://openaccess.thecvf.com/content_cvpr_2018/html/Mahjourian_Unsupervised_Learning_of_CVPR_2018_paper.html but the point of this repo is to provide a consistent baseline reproducing with fidelity the original tensorflow code from the author https://github.com/tinghuiz/SfMLearner

from sfmlearner-pytorch.

@ClementPinard Thanks for your kind answer.

from sfmlearner-pytorch.