search

cvlab-epfl / single-stage-pose

Single-Stage 6D Object Pose Estimation, CVPR 2020
GNU General Public License v3.0
106 stars 17 forks source link

Question about training loss L_p #4

Closed xuhuazhe closed 4 years ago

xuhuazhe commented 4 years ago

In the paper, the loss is \frac{1}{n} \sum ||(\hat{R}p_i + \hat{T} - (Rp_i+T))||. In the code, it is slightly different which is multiplied by intrinsic matrix before computing the distance. Are these two similar or mutually surrogate?

YinlinHu commented 4 years ago

@xuhuazhe In the paper, the error is defined as 3D reconstruction error, in the code, the error is defined as 2D reprojection error. Although 3D reconstruction error can lead to better ADD performance and 2D reprojection error can lead to better REP performance, the difference between them is pretty small in my experience.

xuhuazhe commented 4 years ago

Thank you! Good to know :)