search

ClementPinard / SfmLearner-Pytorch

Pytorch version of SfmLearner from Tinghui Zhou et al.
MIT License
1.01k stars 224 forks source link

Question about the poses predicted by the posenet #153

Closed AyonRRahman closed 11 months ago

AyonRRahman commented 11 months ago

I have a confusion. What are the poses we are getting. Are these to go from the reference image frame to target image frame. So what i mean is, lets say, we got the T=[R, t] from the one of the predictions. if I multiply a 3d point x with it like , Rx+t. Are going from the reference image frame to the target image frame? or the other way from target image frame to reference image frame. Thank you.

ClementPinard commented 11 months ago

hi, it's the other way around, as it can be seen in the code : https://github.com/ClementPinard/SfmLearner-Pytorch/blob/master/inverse_warp.py#L183

if we have a point in the target frame coordinate system, we can get its coordinates in the reference frame system and thus know where it would be projected on the screen (and thus know its color)

so it's not pose per se, but rather the inverse pose.

If you are only interested in training depth, it does not matter, but if you are interested in odometry as well, then you will need to invert the transformation vectors, see how it's done in the evaluation : https://github.com/ClementPinard/SfmLearner-Pytorch/blob/master/test_pose.py#L75

AyonRRahman commented 11 months ago

Thank you. Understood.