sportsbd 0.1.0

sportsbd: a 4-class 3D CNN-based shot boundary detection toolkit for sports videos.

sportsbd

sportsbd: a 4-class 3D CNN-based shot boundary detection toolkit for sports videos.

This library provides a simple Python and command-line interface around a 3D CNN model (r2plus1d_18) for detecting shot boundaries in sports videos. The model predicts four classes:

• hard – hard cuts
• fadein – fade-in boundaries
• logo – logo / overlay-based transitions
• NaN – no boundary

Installation

Prerequisites:

• Python 3.11.14 (recommended, tested and working)
• PyTorch >= 2.3.0 for MPS Conv3D support (PyTorch 2.9.1 recommended)
• TorchVision >= 0.18.0 (TorchVision 0.24.1 recommended)
• NumPy 1.26.4 (tested version)

Recommended setup (tested and working with MPS Conv3D):

# Python 3.11.14 (use pyenv, conda, or your preferred Python version manager)
python --version  # Should show 3.11.14

# Install PyTorch 2.9.1 and TorchVision 0.24.1
pip install torch==2.9.1 torchvision==0.24.1

# Or install latest compatible versions
pip install torch torchvision

# Then install sportsbd
pip install sportsbd

Then install sportsbd:

From PyPI:

pip install sportsbd

Or from source:

Option 1: Install as a package (recommended)

git clone https://github.com/mehdih7/sportsbd.git
cd sportsbd
git lfs install  # Install Git LFS (required for model weights)
git lfs pull     # Download model weights
pip install .

Option 2: Install dependencies only

git clone https://github.com/mehdih7/sportsbd.git
cd sportsbd
git lfs install
git lfs pull
pip install -r requirements.txt
# Then use: python -m sportsbd.cli or add sportsbd/ to PYTHONPATH

For development:

pip install -r requirements-dev.txt
pip install -e .  # Install in editable mode

Quickstart

Python API – run inference on a video

from sportsbd import load_model, run_video_inference, get_available_device

# Load model (auto-detects best device: cuda > mps > cpu)
model = load_model("data/models/best.pt")

# Or explicitly specify device
model = load_model("data/models/best.pt", device="mps")  # Apple Silicon
# model = load_model("data/models/best.pt", device="cuda")  # NVIDIA GPU
# model = load_model("data/models/best.pt", device="cpu")   # CPU

# Run inference on a video (auto-detects device if not specified)
detections = run_video_inference(
    video_path="video.mp4",
    checkpoint_path="data/models/best.pt",
    threshold=0.7,
    stride=4,
    t_frames=16,
    fps=25,
)

# Detections is a list of dicts with:
# - 'frame_idx': frame index
# - 'timestamp_ms': timestamp in milliseconds
# - 'confidence': maximum boundary class probability
# - 'class_probs': per-class probabilities
print(f"Found {len(detections)} shot boundaries")

CLI – run inference on a video

sportsbd infer \
  --video video.mp4 \
  --checkpoint data/models/best.pt \
  --threshold 0.7 \
  --stride 4 \
  --t-frames 16 \
  --fps 25 \
  --out detections.json

Note: The --device option is optional. If not specified, sportsbd automatically detects the best available device (CUDA > MPS > CPU).

Important: Since the r2plus1d_18 model uses 3D convolutions, you should use a PyTorch build with MPS support if you want to run on Apple Silicon GPUs. The configuration below has been tested and confirmed to work with MPS Conv3D:

• Python 3.11.14
• PyTorch 2.9.1
• TorchVision 0.24.1
• NumPy 1.26.4

Supported devices:

• cuda – NVIDIA GPU (Linux/Windows) - supports Conv3D
• mps – Apple Silicon GPU (macOS) - supports Conv3D with PyTorch >= 2.3.0
• cpu – CPU fallback - always works

Model checkpoints

Model weights should be placed in the data/models/ directory. The repository includes a pre-trained model at data/models/best.pt (358MB).

Note: Model weights are tracked with Git LFS due to their size. After cloning, run:

git lfs install
git lfs pull

sportsbd expects a PyTorch checkpoint (.pt/.pth) containing at least:

• state_dict – model weights compatible with torchvision.models.video.r2plus1d_18
• config – a dictionary with at least:
  • MODEL_NAME – model architecture name (defaults to "r2plus1d_18")
  • NUM_CLASSES – number of output classes (defaults to 4)

You can use any path to your checkpoint file - the examples use data/models/best.pt as a convention.

Development

To run tests:

pip install -e ".[dev]"
pytest