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

3D Pose Augmentation #5

Open Anshul-Gupta24 opened 4 years ago

Anshul-Gupta24 commented 4 years ago

Hi, I am trying to merge the PVNet code with your single stage code. As part of training, the images undergo augmentation such as rotation, cropping and scaling. I have applied the corresponding augmentation to the 3D pose but the network is very slow to train. After 120 epochs, it is performing very poorly even on train data. Removing augmentation altogether causes the network to overfit. It would be great if you could suggest where I am going wrong. Thanks!

magicknight commented 4 years ago

rotation and scaling? why?

linharrrrrt commented 4 years ago

rotation and scaling? why?

the original PVNet code do those things to augment the data.

pra-dan commented 3 years ago

Hey @Anshul-Gupta24 How did u make the end-to-end pipeline to use this repo + PVNet to get 6 DoF ?

I am running it on this colab notebook but cannot understand how to get the 6 DoF. 😢

Yipingggg commented 1 year ago

Hi, I am trying to merge the PVNet code with your single stage code. As part of training, the images undergo augmentation such as rotation, cropping and scaling. I have applied the corresponding augmentation to the 3D pose but the network is very slow to train. After 120 epochs, it is performing very poorly even on train data. Removing augmentation altogether causes the network to overfit. It would be great if you could suggest where I am going wrong. Thanks!

Have you sorted out the problem in the end? I think you should correct the ground truth pose after the augmentation.