search

aau-cns / poet

PoET: Pose Estimation Transformer for Single-View, Multi-Object 6D Pose Estimation
Other
71 stars 12 forks source link

When training the lmo dataset, all results are 0, and the translation and rotation errors are large. #27

Open Fusica opened 5 days ago

Fusica commented 5 days ago

I encountered several issues when trying to train on the lom dataset and reproduce the author's results:

  1. First, I found that in lmo2poet.py, the division method for the lmo dataset is quite different from what I downloaded from bop. The lmo downloaded from bop does not have the train_pbr and train_synt subfolders. I would like to ask which version of lmo the author used.

  2. Then, when I loaded lmo_maskrcnn_checkpoint.pth.tar during training, the ADD was all 0. I want to know if it's because I made a mistake in the dataset generation step, or if there is something wrong with my training hyperparameters. I have referred to some of the hyperparameter settings provided in the author's link, but the results have not improved at all.

I sincerely hope the author can point out where my problem lies, as I am very eager to follow your work.

Fusica commented 5 days ago

I tried using the command python main.py --epochs 50 --batch_size 24 --output_dir outputs/202410231444 --resume weights/poet_lmo_maskrcnn.pth --eval to test the pth file you provided, and I got the following results. Does this match your results, and does this result prove that there is no problem with my dataset?

image image image image image

The angle error feels large, with 40°.