OpenSportsLib

OpenSportsLib is a configuration-driven Python library for sports-video experiments. The currently runnable task APIs are classification, localization (action spotting), and visual question answering (VQA).

Canonical YAML config -> task wrapper -> dataset and model builders
-> trainer, inferer, or evaluator -> task metric / OSL JSON prediction payload

Retrieval and captioning are roadmap areas. They do not currently have runnable task wrappers, dataset dispatch, or model routes in this package.

Start here

1. Install OpenSportsLib.
2. Read the first workflow. The package does not ship a demo dataset, so you must supply media and manifests.
3. Choose a canonical configuration and prepare an OSL JSON split.
4. Train, infer, evaluate, and explicitly save results using the workflow guide.

Guides by audience

• New users: installation, project structure, and first workflow.
• Researchers: configuration, data format, model zoo, and workflows.
• Developers: architecture, supported extensions, and the configuration developer guide.
• Contributors: contributing.
• Remote serving: inference server.

License

OpenSportsLib is available under AGPL-3.0 and commercial licensing terms. See the repository license files for authoritative terms.

Acknowledgments

OpenSportsLib is developed within the broader OpenSportsLab effort for sports video understanding. Core contributors affiliated with KAUST include:

• Jeet Vora — Remote Research Engineer
• Dr. Merey Ramazanova — Post-Doc
• Dr. Silvio Giancola — Research Scientist