Open WHHHHY opened 6 days ago
Thanks for your interest of our work! I assume that you would like to use a single network to predict multiple objects simultaneously, like our Linemod experiments reported in the paper. For this dataset we have separate files for dataset loading , training and testing .
Dataset codes are mainly in checkerpose/tools_for_LM/get_lm_datasets.py
and checkerpose/lm_dataset_pytorch.py
. The dataset paths can be modified in checkerpose/tools_for_LM/get_lm_datasets.py
.
Training codes can be found in checkerpose/pretrain_lm.py
and checkerpose/train_lm.py
. (The name of pretrain
is kind of misleading, it actually means training only the low level layers of the whole network.) And test file is checkerpose/test_lm.py
. For these files, reset the dataset related codes.
Also, write new configs similar to checkerpose/config/lm/init_gnn2_hrnetw18_npt512_lm.txt
(for pretrain) and checkerpose/config/lm/hr18GNN2_res6_gnn3Skip_mlpQuery_lm.txt
(for train).
For Linemod experiments, we use two training subsets. One is the images from BOP website, and the other one is a synthetic dataset downloaded from GDR-Net GitHub repo. You may need to write
training_data_folder_2 = none
in your config files if you only have one single training subset.
Also, in our default setting, we randomly reset the background images
https://github.com/RuyiLian/CheckerPose/blob/fb725bed2a9eb6c3646c476c017d9ad00aa63a28/checkerpose/config/lm/hr18GNN2_res6_gnn3Skip_mlpQuery_lm.txt#L11
We use VOC 2012
as background images. The details can be found in our README
(data preparation, 2 and 4)
You can just discard codes of changing background if it is unnecessary.
Thanks for your excellent work! And if I want train my customized datasets made in linemod format,how should I modify the paths and codes?