FoundationPose_IS

Follow these steps to run FoundationPose on the IndustryShapes dataset.

Apply some minor modifications to the FoundationPose repo:

cd FoundationPose
git apply ../diff.patch

Data preparation & Docker container setup

Follow the Data Preparation steps in the FoundationPose directory.

Download the IndustryShapes dataset from here and place it in the FoundationPose/BOP directory.

For the model-based setup:

Run:

export BOP_DIR = /path/to/FoundationPose/BOP

And then:

python run_linemod.py --dataset_dir /path/to/FoundationPose/BOP/IndustryShapes --use_reconstructed_mesh 0

For the model-free setup:

You can run BundleSDF, as the authors of FoundationPose suggest, on the IndustryShapes onboarding sequences (up) to get the object pose relative to the camera for each frame. These can then be used by FoundationPose as reference views to reconstruct the mesh.

Run:

python bundlesdf/run_nerf.py --ref_view_dir /path/to/ref_views --dataset lmo

And then:

python run_linemod.py --dataset_dir /path/to/FoundationPose/BOP/IndustryShapes --use_reconstructed_mesh 1 --ref_view_dir /path/to/ref_views

Evaluation

FoundationPose creates a yml file in the /debug dir. You can run

cd ..
python eval_bop.py

to get a BOP format prediction csv. You can then use bop_toolkit for the evaluation.