The ./toolkit folder contains scripts to prepare data.
Download the dataset from https://bop.felk.cvut.cz/datasets/.
More specifically, only All test images of LM (Linemod) has to be downloaded.
(Only the test folder contains real images which are used for training and testing in previous works, including ours)
Extract the test files to folder $(DeepIM_root)/data/LINEMOD_6D/LM6d_origin/test:
unzip path/to/lm_test_all.zip -d $(DeepIM_root)/data/LINEMOD_6D/LM6d_origin/
Run these commands successively to prepare LM6d_refine:
Our processed models (models.tar.gz), train/val split (LINEMOD_6D_image_set.tar.gz) and PoseCNN's results (PoseCNN_LINEMOD_6D_results.tar.gz) can be found on Google Drive
Download and extract them in folder$(DeepIM_root)/data/LINEMOD_6D/LM6d_converted/LM6d_refine
which shall like:
LM6d_refine/models/ape, benchvise, ...
LM6d_refine/image_set/observed/ape_all.txt, ...
LM6d_refine/PoseCNN_LINEMOD_6D_results/ape, ...
After putting all the files in correct location, you can just run
sh prepare_data.sh
to prepare original dataset and synthetic data for LINEMOD.
We use indoor images from Pascal VOC 2012 (download link) as the background of these synthetic during training.
Download and extract it in the $(DeepIM root)/data, which will like $(DeepIM_root)/data/VOCdevkit/VOC2012.
Support files for other dataset will be released later.