Accurate baseball player pose refinement using motion prior guidance

---

Contributions

• We propose Baseball Player Pose Corrector (BPPC), an optimization technique for refining keypoints in baseball batting, leveraging prior knowledge of the 3D swing motion.

• We introduce a 4D keypoint projection method that accurately matches 3D standard motions to 2D test videos, regardless of the differences between the standard motion and test swing videos.

• We propose a loss function that adaptively optimizes poses based on keypoint confidence and movements.

• BPPC improves the quantitative and qualitative performance of state-of-the-art HPE models on benchmark datasets.

Project Structure

BPPC/
├── core/
│   └── model.py                  # BPPC model (optimization module)
├── runners/
│   ├── opt/opt_bppc.py           # Step 1: Run BPPC optimization, save results
│   └── eval/eval_bppc.py         # Step 2: Evaluate & visualize saved results
├── scripts/
│   ├── run_opt_bppc.sh           # Run optimization for all test cases
│   ├── run_eval_bppc.sh          # Run evaluation for all test cases
│   ├── run_one_bppc.sh           # Run a single test case (opt + eval)
│   └── run_pipeline.sh           # Full pipeline (opt → eval → Excel export)
├── utils/
│   ├── accuracy_bppc.py          # Accuracy calculation utilities
│   └── visualize_bppc.py         # Visualization utilities
├── lib/                          # Backbone models and dataset utilities
├── grid_sample1d/                # Custom 1D grid sampling CUDA extension
├── data/
│   ├── images/                   # Input images ({left,right}_final/)
│   ├── frozen/                   # Pre-extracted 2D keypoints & scores
│   ├── gt_2D/                    # Ground truth 2D keypoints
│   └── sm_3D/                    # 3D standard motion data
├── demo/
│   ├── bppc/                     # BPPC optimization results (.npz)
│   └── output/                   # Evaluation results & visualizations
└── aggregate_results.py          # Collect all results into Excel

Getting Started

Environment Requirement

Clone the repo:

git clone https://github.com/OSHMOS/BPPC.git

Install the bppc requirements using conda and pip:

conda create -n bppc python=3.12 -y

# Install PyTorch
# The following command is an example and should be modified according to your CUDA version and system environment
pip install torch==2.4.1 torchvision==0.19.1 torchaudio==2.4.1 --index-url https://download.pytorch.org/whl/cu118

cd grid_sample1d/
python setup.py install
cd ..

pip install opencv-python tabulate scipy tqdm yacs numba scikit-image filterpy

Prepare the checkpoints:

${POSE_ROOT}
    `-- lib
        `-- checkpoint
            |-- darkpose
            |   |-- w32_384×288.pth
            |   `-- w48_384×288.pth
            |-- hrnet
            |   |-- pose_hrnet_w32_384x288.pth
            |   `-- pose_hrnet_w48_384x288.pth
            |-- resnet
            |    |-- pose_resnet_50_384x288.pth
            |    |-- pose_resnet_101_384x288.pth
            |    `-- pose_resnet_152_384x288.pth
            `-- yolo3.weights

Test

The pipeline is split into two stages: Optimization and Evaluation.
Pre-extracted 2D keypoints from baseline models are stored in data/frozen/, so backbone inference does not need to be re-run.

Run All Test Cases (Full Pipeline)

conda activate bppc

# Step 1: Optimization (saves results to demo/bppc/)
bash scripts/run_opt_bppc.sh

# Step 2: Evaluation + Visualization (reads from demo/bppc/, saves to demo/output/)
bash scripts/run_eval_bppc.sh

# Or run both steps + Excel export at once
bash scripts/run_pipeline.sh

Run a Single Test Case

Useful for quick debugging or per-sample inspection:

# bash scripts/run_one_bppc.sh <handed> <folder_number> [model_name]

# Example: left-handed batter, folder 0170, ResNet-152
bash scripts/run_one_bppc.sh left 0170 res152

# Example: all models for one folder
bash scripts/run_one_bppc.sh left 0170

# Handed options: left / right
# Model options:  res50, res101, res152, hw32, hw48, darkw32, darkw48, all

Collect Results into Excel

python aggregate_results.py
# Output: eval_bppc_results.xlsx (Body Results + Conf Results sheets)

Citation

@article{OH2025,
    title = {Accurate baseball player pose refinement using motion prior guidance},
    journal = {ICT Express},
    year = {2025},
    issn = {2405-9595},
    doi = {https://doi.org/10.1016/j.icte.2025.03.008},
    url = {https://www.sciencedirect.com/science/article/pii/S2405959525000360},
    author = {Seunghyun Oh and Heewon Kim},
    keywords = {Human pose estimation, Human pose refinement, Deep learning}
}

Acknowledgement

• The repo (grid_sample1d) is based on Grid Sample 1d. Thanks for their well-organized code!