How ispgroupucl/deepsportlab works
Project type and purpose
• DeepSportLab is a unified, full-stack machine-learning framework focusing on three tasks in team sports scenes: ball detection, player instance segmentation, and pose estimation.
• It provides both training and inference pipelines (Python CLI, notebooks) and exposes C++ examples for embedding/integration.
File structure
• Top-level:
• openpifpaf/ ← Core Python package
• cpp/ ← C++ examples using the library
• docs/, guide/ ← Documentation and tutorials
• tests/ ← Unit and integration tests
• predict.py, train.py ← CLI entry points for inference and training
• export_onnx.py ← Model export for production
• openpifpaf/ sub-packages:
• datasets/ ← Data loading and collate functions
• transforms/ ← Preprocessing pipelines
• encoder/ decoder/ ← Task-specific encoding (part intensity fields) and decoding (spatial embeddings)
• network/ ← Model definition, loss computation, training loop
• visualizer/ show/ ← Output visualization and CLI helpers
• Configuration & tooling:
• setup.py, setup.cfg ← Packaging
• requirements.txt ← Python dependencies
• CI/CD workflows (.github/, .travis.yml)
• notebooks (.ipynb) ← EDA, guide examples
README for additional insights
• Describes the unified single-model approach combining part intensity fields and spatial embeddings.
• Mentions evaluation on the DeepSport basketball dataset, export options, and licensing.