This repository provides an end-to-end pipeline for:

• Preparing data for YOLO-based object detection using Ultralytics’ YOLO.
• Training a YOLO model to detect a specific object.
• Training a simple pose estimation model (SimplePoseNet) on BOP-format data.

bpc/
  yolo/
    prepare_data.py   # Prepare data in YOLO format
    train.py          # Train YOLO using Ultralytics
    configs/         # YOLO configuration files (.yaml)
    models/          # YOLO trained weights (.pt)

  pose/
    checkpoints/     # Stores pose model checkpoints
    models/          # Network definitions, losses, etc.
    trainers/        # Training logic for pose estimation

  utils/
    data_utils.py    # BOP dataset loading + transforms

  inference/
    # Scripts for epipolar matching, YOLO detection, and match pipelines

train_pose.py       # Pose estimation training script
datasets/           # Contains BOP and YOLO data
runs/               # YOLO training runs, logs, etc.

To set up the environment, follow these steps (tested on Ubuntu with an NVIDIA GPU). The environment name is bop.

cd docker/
docker build . -t bpc:2025.1.31

docker run -p 8888:8888 --shm-size=24g --runtime nvidia --gpus all -v $(pwd):/code -ti bpc:2025.1.31 bash
cd /code

Hint: Ensure you return to the bpc_baseline folder before running the pipeline. Change shm-size with what fits your needs, e.g. 8g, 16g, etc.

bash download_data.sh

Note: If you plan to download the dataset and train your model, remember that for submission you should delete the downloaded datasets in the datasets folder. This prevents the Docker image from growing too large.

Convert BOP data to YOLO format:

python3 bpc/yolo/prepare_data.py \
    --dataset_path "datasets/ipd/train_pbr" \
    --output_path "datasets/yolo11/train_obj_8" \
    --obj_id 8

Hint: You can modify the visibility threshold (visib_fract) in prepare_data.py.

python3 bpc/yolo/train.py \
    --obj_id 8 \
    --data_path "bpc/yolo/configs/data_obj_8.yaml" \
    --epochs 20 \
    --imgsz 1280 \
    --batch 16 \
    --task detection

python3 train_pose.py \
  --root_dir datasets/ipd \
  --target_obj_id 8 \
  --epochs 50 \
  --batch_size 32 \
  --lr 1e-4 \
  --num_workers 8 \
  --checkpoints_dir bpc/pose/pose_checkpoints/ \
  --loss_type quat

Hints:

• Use --use_real_val to utilize the val folder in datasets/ipd/ as the validation set. Without this flag, the script will split the train_pbr data in an 80:20 ratio.
• Use --resume to continue training from the last checkpoint if available.

wget https://huggingface.co/DrQY/BOP_baseline_model/resolve/main/models.zip
unzip models.zip
rm models.zip

jupyter notebook --ip=0.0.0.0 --allow-root --port=8888
# Go to localhost:8888 on your browser
# Run "Inference Notebook.ipynb"

• Ensure CUDA 12.1 drivers are installed and PyTorch recognizes the GPU (nvidia-smi).
• BOP dataset must follow standard conventions (train_pbr, test, etc.).
• Update bpc/yolo/configs/data_obj_8.yaml with the correct dataset paths.
• If encountering module import errors, try:

  python -m idp_codebase.pose.train ...

  or add __init__.py files where necessary.

For detailed documentation, please refer to this page.