Awesome

6Dpose

Welcome to the official repository of Revisiting Fully Convolutional Geometric Features for Object 6D Pose Estimation, or FCGF6D, to be presented at Recovering 6D Pose Workshop at ICCV 2023. The repository is still work in progress as I am simplifing and refactoring the code, but everything is here. I am currently busy finishing my Master's Degree, but I hope to get all the code done by end of October. Feel free to contact me if you have any questions: jaime.corsetti98@gmail.com.

TODOS:

• Provide checkpoints
• Instructions to generate YCBV training set
• Add metadata (e.g. object splits)

Setup

Our work is based on Minkoski Engine and implemented with Pytorch Lightning. Please install them following the official instructions before proceeding. We strongly suggest to use a virtual environment for the installation. After this you can install all the other libraries in the requirements.txt file.

Preprocessing

Download the YCBV test split and the OccludedLinemod (LMO) test and train_pbr split from the BOP challenge website. All contents of YCBV should go in a data_ycbv folder, all contents of LMO should go in a data_lmo folder.

After this, generate the ground truth by running:

python utils_scripts/make_gt.py --dataset lmo --split test
python utils_scripts/make_gt.py --dataset ycbv --split test

and place the resulting files in the eval_gts folder. This formats the ground truth in an easy-to-read file to be used for the compute_metrics.py script.

Generate the segmentation mask in a better format by running:

python utils_scripts/preprocessing_mask.py --dataset lmo --split test
python utils_scripts/preprocessing_mask.py --dataset ycbv --split test

This will turn the N binary masks for each image provided by BOP in a single mask, which speeds up the I/O processes.

Finally apply the hole filling algorithms:

python utils_scripts/hole_filling.py --dataset lmo --split test
python utils_scripts/hole_filling.py --dataset ycbv --split test

to generate the hole filled depth masks as in PVN3D and FFB6D. Hole filling algorithm and mask preprocessing should be carried out also for any training splits.

Train

For OccludedLinemod and YCBV trainings

CUDA_VISIBLE_DEVICES=0 python run_train.py --exp exp_lmo --dataset lmo --split_train 'train_pbr' --split_val 'test' --bs 8 --n_epochs 10 --freq_valid 2 --freq_save 2  --augs_erase True --augs_rgb True --depth_points 50000 --arch mink34
CUDA_VISIBLE_DEVICES=0 python run_train.py --exp exp_ycbv --dataset ycbv --split_train 'custom_train' --split_val 'test' --bs 8 --n_epochs 110 --freq_valid 10 --freq_save 10  --augs_erase True --augs_rgb True --depth_points 20000 --arch mink50

please see utils/parsing.py (parse_train_args function) for the complete list of options at training time.

Test

CUDA_VISIBLE_DEVICES=0 python --exp exp_lmo --depth_points 50000 --arch mink34 --checkpoint 'epoch=0009' --split test --solver teaser
CUDA_VISIBLE_DEVICES=0 python --exp exp_ycbv --depth_points 20000 --arch mink50 --checkpoint 'epoch=0109' --split test --solver ransac

please see utils/parsing.py (parse_test_args function) for the complete list of options at evaluation time.

The above scripts will produce the poses, in order to compute the metrics also the following is necessary:

Metrics

python compute_metrics.py --exp 'exp_lmo' --split 'test' --checkpoint 'epoch=0009' --solver teaser 
python compute_metrics.py --exp 'exp_ycbv' --split 'test' --checkpoint 'epoch=0109' --solver ransac 

Acknowledgements

Parts of this work was based on the code provided by PVN3D (ADDS-AUC metric code) and Fast Depth Completition for the depth hole-filling algorithm.

Citation

Please cite FCGF6D if you use this repository in your publications:

@inproceedings{corsetti2023fcgf6d,
  author = {Corsetti, Jaime and Boscaini, Davide and Poiesi, Fabio},
  title = {Revisiting Fully Convolutional Geometric Features for Object 6D Pose Estimation},
  booktitle = {International Workshop on Recovering 6D Object Pose (R6D)},
  year = {2023}
}