This repository implements the methods described in Physical Plausibility of 6D Pose Estimates in Scenes of Static Rigid Objects. The code and data required to reproduce the results in Figures 5, 7 and 8 are provided.

The code has been tested on Ubuntu 16.04 with Python 3.6. To set-up the Python environment, use Anaconda and the provided YAML file:

conda env create -f environment.yml --name plausibility

conda activate plausibility

We use the pose-error functions and rendering implemented in the BOP toolkit. Please make sure that the BOP toolkit submodule is correctly initialized and see this repository for installation instructions. Finally, replace the file bop_toolkit_lib/renderer_py.py with src/renderer_py.py. It patches the Python-based renderer in the BOP toolkit, allowing to apply a transformation when loading objects.

The object models from the YCB VIDEO dataset can be downloaded on the BOP Challenge 2020 page. Please change YCBV_PATH in src/evaluation.py to reflect your model directory.

Additional meta information for the models and precomputed samples for evaluation are provided in data.

Running python src/evaluation.py --target [TARGET] --mode [MODE] generates pose estimates for evaluation, computes the baseline pose-error functions (MSSD, MSPD, VSD) as well as our pose errors. The results are visualized in an animated plot. TARGET selects the artifical scene to evaluate: bowl for the scene in Figure 5, marker for Figure 7 and clamp for Figure 8 in the paper. MODE selects whether to apply rotation or translation error to the ground-truth pose.

If you use this repository in your publications, please cite

@article{bauer2020plausibility,
    title={Physical Plausibility of 6D Pose Estimates in Scenes of Static Rigid Objects},
    author={Bauer, Dominik and Patten, Timothy and Vincze, Markus},
    booktitle={European Conference on Computer Vision Workshops (ECCVW)},
    year={2020}
}