Steps to run Experiments

Preprocessing the data

Run the following command locally (not inside Docker):

DATASET_PATH=/path/to/IndustryShapes
python3 utils/prepare_dataset.py \
  --data_folder $DATASET_PATH/train/ \
  --output_folder data/preprocessed \
  --models_path $DATASET_PATH/models_eval \
  --obj_map  obj_names.json\
  --scenes "000001 000002 000003 000004 000005 000006 000007 000008 000009 000010 000011 000012 000013"

Build the docker image

From the directory containing the Dockerfile.noetic (docker/), run:

For training:

cd docker
docker build -t nvidia-dope:train -f train/Dockerfile.noetic ..

For inference:

cd docker
docker build -t nvidia-dope:test -f train/Dockerfile.noetic ..

Training the model

Apply some minor modifications to Deep_Object_Pose repo:

cd Deep_Object_Pose
git apply ../diff.patch

Activate docker container

./run_dope_docker.sh dope-training nvidia-dope:train

From the project root (outside the Docker directory), run:

python3 -m torch.distributed.launch --nproc_per_node=1 train.py --network dope --epochs 100 --batchsize 1 --outf tmp/ --data /path/to/data/preprocessed

Running inference

Activate docker container

./run_dope_docker.sh dope-testing nvidia-dope:test

Example for running inference on one test scene:

python3 inference.py --config /path/to/config/test_scenes/config_pose_000001.yaml --camera /path/to/config/test_scenes/camera_info_000001.yaml --data /path/to/data/000001/rgb/

Example running on all test scenes

for i in $(seq -w 1 8); do
  python3 inference.py --config /path/to/config/test_scenes/config_pose_00000$i.yaml \
                       --camera /path/to/config/test_scenes/camera_info_00000$i.yaml \
                       --data /path/to/test-dataset/00000$i/rgb/ \
                       --outf /path/to/output_experiments/00000$i
done

Evaluation using bop toolkit

Create bop_predictions.csv file using the following script

python utils/results_to_bop.py \
  --root_folder /path/to/output_experiments/ \
  --outf /path/to/evaluation_results/ \
  --modelpath /path/to/IndustryShapes/models_eval

(Optional) If you want to visualize the results, add --visualize flag to the command above.