GDRNPP Refinement

The refinement module of GDRNPP.

Environment

Create a new anaconda environment using the provided .yaml file.

conda env create -f environment.yaml
conda activate refine

Preparing Data

Link the datasets in the following folders with symlink.

├── local_data
    ├── VOCdevkit
        ├── VOC2012
    ├── bop_datasets
        ├── hb
        ├── icbin
        ├── itodd
        ├── lm
        ├── lmo
        ├── tless
        ├── tudl
        ├── ycbv

The pretrained models are available on onedrive (password: groupji).

Training

We train a model separately for each of the BOP datasets. To train the TUD-L refinement model, run

python train.py --dataset tudl --batch_size 12 --num_inner_loops 10 --num_solver_steps 3 [--pbr_only]

To train a refinement model on other BOP datasets, replace tudl with one of the following: tless, lmo, hb, ycbv, icbin, itodd

Testing

1. Generate GDRN results with save_gdrn.sh and put the trained RAFT models in folder "model_weights".

2. Run

./test_fast.sh <DATASET> <RESULT_PATH> <GPU_ID>

Use "test_fast_pbr.sh" for PBR results of YCB-V, T-LESS and TUD-L.

3. Run

python -m additional_scripts.convert_result --tar_dir <TAR_DIR> --dataset <DATASET>

Acknowledgement

This repository makes extensive use of code from the Coupled-Iterative-Refinement and Cosypose Github repository. We thank the authors for open sourcing their implementation.